Multi-modal data fusion method and system for extreme strong wind of power transmission line

Abstract

The invention discloses a multi-modal data fusion method and system for extreme strong wind of a power transmission line, and the method comprises the steps: collecting strong wind meteorological data, remote sensing images and data of different modals of topographic features, carrying out the projection superposition, forming a standardized data cube, and carrying out the zero filling and three-dimensional convolution operation, thereby achieving the fusion of the extreme strong wind of the power transmission line. Each data cube is divided into space-time image block cubes, and the space-time image block cubes are converted into feature unit sequences through linear transformation; and taking the network coordinates of the data blocks as indexes, adding a position code for each feature sequence, and generating a final meteorological, remote sensing and topographic feature unit sequence. Then, the model extracts common flow features representing a shared background and unique flow features reflecting data source features from the sequences respectively, and finally, intelligent fusion is performed by taking the common flow features as key values and the unique flow features as queries, so that multi-modal data fusion is realized, and the reliability of a data fusion result is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power system disaster prevention technology, and in particular to a multimodal data fusion method and system for extreme winds on transmission lines. Background Technology

[0002] Previous methods for fusing multi-source wind data include traditional methods and machine learning-based methods. Traditional wind data fusion methods primarily use optimal interpolation or 3D variational methods to fuse scatterometer sea surface wind fields with ground meteorological station observations, outputting 0.25°–0.5° grid products. Alternatively, shallow models such as random forests, extreme gradient boosting, and supervised learning algorithms are used to perform nonlinear regression on satellite cloud top brightness temperature, radar reflectivity, climate reanalysis datasets, and station gusts to identify thunderstorm winds with a hit rate of approximately 80%. These methods can effectively fuse the features of multi-source wind data. However, when dealing with mountainous terrain, due to the large elevation differences and low station density, fusion is mainly performed by directly "weighted averaging" or "stitching," ignoring the unique and shared features of the multi-source data. This leads to the loss of some information in the data fusion results, causing the fusion results to deviate from the actual situation and reducing reliability. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a multimodal data fusion method and system for extreme winds on transmission lines, thereby solving the technical problem of low reliability of data fusion results obtained by existing multimodal data fusion methods for extreme winds on transmission lines.

[0004] A first aspect of this invention provides a multimodal data fusion method for extreme winds on transmission lines, the method comprising: Acquire gale meteorological data, remote sensing data, and topographic data of the area to be monitored; Meteorological data, remote sensing data, and topographic data are projected and overlaid to obtain their respective data cubes; Zero-padding and convolution segmentation operations are performed on each data cube to obtain the spatiotemporal image block cube corresponding to each data cube. A linear transformation is performed on each spatiotemporal image block cube to obtain the corresponding feature unit sequence; Using the network coordinates of each data block in each spatiotemporal image block cube as the spatiotemporal index, positional encoding is added to the corresponding feature unit sequence to obtain the corresponding final feature unit sequence. The corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence. Based on the obtained final feature unit sequences, unique flow extraction and common flow extraction are performed respectively to obtain common flow features and unique flow features; The common flow features are used as keys and the unique flow features are used as queries to fuse the data and obtain the fusion result.

[0005] In one possible implementation of the first aspect, when obtaining the corresponding final feature unit sequence, the method further includes: By projecting terrain dynamic parameters onto the terrain data cubes in each data cube, higher-order terrain dynamic features are obtained. Higher-order terrain dynamic features are added to the corresponding feature unit sequence to obtain the final terrain feature unit sequence.

[0006] In one possible implementation of the first aspect, based on the obtained final feature unit sequences, unique flow extraction and common flow extraction are performed respectively to obtain common flow features and unique flow features, including: Unique flow extraction is performed on the final remote sensing feature unit sequence to obtain the dynamic evolution characteristics of cloud clusters; Unique flow extraction is performed on the final terrain feature unit sequence to obtain terrain features; The meteorological features are obtained by extracting unique flows from the final meteorological feature unit sequence. The dynamic evolution characteristics, topographic features, and meteorological features of cloud clusters were identified as unique flow characteristics. Shared flow features are extracted based on the final remote sensing feature unit sequence, the final topographic feature unit sequence, and the final meteorological feature unit sequence.

[0007] In one possible implementation of the first aspect, unique flow extraction is performed on the final remote sensing feature unit sequence to obtain the dynamic evolution features of the cloud cluster, including: Feature analysis is performed on the final remote sensing feature unit sequence to obtain a new terrain feature unit sequence; The dynamic evolution patterns of cloud clusters are extracted from the sequence of new terrain feature units to obtain the dynamic evolution characteristics of cloud clusters.

[0008] In one possible implementation of the first aspect, a unique flow extraction is performed on the final sequence of terrain feature units to obtain terrain features, including: The final sequence of terrain feature units is deconvolved to obtain the first spatial map; The slope component is obtained by performing slope convolution on the first spatial graph in multiple directions using multiple direction kernels. By concatenating the slope component and the Laplace elevation residual, a two-dimensional physical vector of the terrain is obtained. Terrain features are obtained by convolving and projecting two-dimensional physical vectors.

[0009] In one possible implementation of the first aspect, unique flow extraction is performed on the final meteorological feature unit sequence to obtain meteorological features, including: The final meteorological feature unit sequence is deconvolved to obtain the second spatial map; Based on the second spatial map, the meteorological two-dimensional physical vector is obtained; The attention map is obtained by extracting gust jump features based on meteorological two-dimensional physical vectors. The attention map is convolved and projected to obtain meteorological features.

[0010] In one possible implementation of the first aspect, common flow features are extracted based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain common flow features, including: Spatiotemporal skeletons were extracted from the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain coordinate spatiotemporal skeletons. Based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence, cross-source covariance guided weighting operations are performed to obtain gated weights; The common guiding signal is obtained by multiplying the gating weights and the coordinate spatiotemporal skeleton element by element. The shared guiding signal is identified as a shared flow characteristic.

[0011] To address the same technical problem, a second aspect of this invention provides a multimodal data fusion system for extreme winds on transmission lines, comprising: The acquisition module is used to acquire wind meteorological data, remote sensing data, and topographic data of the area to be monitored; The projection overlay module is used to project and overlay meteorological data, remote sensing data and terrain data respectively to obtain their respective data cubes; The convolution segmentation module is used to perform zero-padding and convolution segmentation operations on each data cube to obtain the spatiotemporal image block cube corresponding to each data cube. The linear transformation module is used to perform linear transformations on each spatiotemporal image block cube to obtain the corresponding feature unit sequence. The location encoding module is used to add location encoding to the corresponding feature unit sequence using the network coordinates of each data block in each spatiotemporal image block cube as the spatiotemporal index, so as to obtain the corresponding final feature unit sequence. The corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence. The feature extraction module is used to extract unique flow and common flow based on the obtained final feature unit sequences, to obtain common flow features and unique flow features. The fusion module is used to fuse common flow features as keys and unique flow features as queries to obtain fusion results.

[0012] A third aspect of the present invention provides a computer device, comprising: Memory, used to store computer programs; A processor is used to execute computer programs to implement steps of a multimodal data fusion method for extreme winds on transmission lines, as described in the first aspect.

[0013] A fourth aspect of the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the multimodal data fusion method for extreme winds on transmission lines as described in the first aspect.

[0014] The technical solution of this invention has the following advantages: The multimodal data fusion method for extreme winds affecting power transmission lines provided in this invention first collects data from different modalities of wind meteorology, remote sensing imagery, and terrain features, and then projects and overlays these data to form a standardized data cube. Next, through zero-padding and 3D convolution operations, each data cube is segmented into spatiotemporal image block cubes, which are then linearly transformed into a sequence of feature units. Using the network coordinates of the data blocks as indices, positional encoding is added to each feature sequence to generate the final sequences of meteorological, remote sensing, and terrain feature units. Subsequently, the model extracts common flow features representing shared backgrounds and unique flow features reflecting the characteristics of the data source from these sequences. Finally, intelligent fusion is performed using common flow features as keys and unique flow features as queries, achieving multimodal data fusion and effectively improving the reliability of the data fusion results. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a multimodal data fusion method for extreme winds on transmission lines, as described in an embodiment of the present invention. Figure 2 This is a technical roadmap for the multimodal data fusion method for extreme winds on transmission lines, as described in this invention. Figure 3 The diagram shows the extreme value jump self-attention network module and the SE_Block channel attention network architecture of the multimodal data fusion method for extreme winds on transmission lines in this embodiment of the invention. Figure 4This is a diagram of the Cross-Mamba feature fusion network architecture for a multimodal data fusion method for extreme winds on transmission lines, as described in this embodiment of the invention. Figure 5 This is a system block diagram of a multimodal data fusion system for extreme winds on power transmission lines, as described in an embodiment of the present invention. Figure reference numerals: 500, Multimodal data fusion system for extreme winds on transmission lines; 501, Acquisition module; 502, Projection overlay module; 503, Convolution segmentation module; 504, Linear transformation module; 505, Position encoding module; 506, Feature extraction module; 507, Fusion module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] The multimodal data fusion method for extreme winds on transmission lines provided in this invention embodiment, such as... Figure 1 As shown, Figure 1 The flowchart of the multimodal data fusion method for extreme winds on transmission lines is shown below, including steps S101 to S107. The specific steps are as follows: S101. Obtain meteorological data, remote sensing data and topographic data of the area to be monitored.

[0020] In this embodiment, gale meteorological data, remote sensing data, and topographic data of the monitored area are acquired through ground-based meteorological stations and meteorological satellites. For example, ground-based meteorological stations provide near-surface true wind speed sequences at a sampling frequency of minutes to obtain gale meteorological data. Remote sensing data is obtained by using geostationary meteorological satellites to capture convective cloud evolution features such as bow echoes and backflow gaps in real time through 1km / 10min multispectral imaging, providing upper-air observation basis for subsequent models to identify gale generation and movement paths in advance. High-resolution DEM (Digital Elevation Model) data quantifies topographic parameters such as canyon depth-to-width ratio, ridge shielding angle, and over-mountain slope, providing information on ground undulation. As an example of this embodiment, gust data from 287 meteorological stations in a canyon section in Southwest China were selected, spanning the past five years of thunderstorm seasons (June–September), collecting approximately 1.8 × 10⁻⁶ gusts. 7Record a 1-minute instantaneous gust. Then remove sensor fault codes and negative instantaneous wind speed readings, ensuring the gust value is within 0-50ms. -1 Between these steps, a 15-minute window is used to take the maximum value that slides for 3 seconds to generate a 15-minute gust sequence. At this point, the gust values ​​in CSV (Comma-Separated Values) format are converted into a table to obtain the gale meteorological data.

[0021] We acquired a geostationary satellite image with a raw spatial resolution of 4 km and a visible light band resolution of 0.5 km, with a temporal resolution of 10 min. First, we used super-resolution reconstruction to divide it into 1 km segments. Then, we took one image every 15 min, also cut into 1024×1024 frames, and selected 8 convection-sensitive bands with wavelengths of 0.47-13.5 μm. Each frame was a standardized remote sensing image, thus obtaining remote sensing data.

[0022] A global elevation dataset with a unified format, coordinate system, and elevation benchmark was acquired, with a resolution of 30m. The 30m elevation maps were then enlarged / merged into 1km topographic base maps covering the same geographical area, with original elevations ranging from 143 to 7556m. Z-score normalization yielded a 1024×1024×6 topographic base map. The values ​​in the base map included aspect ratio, shading angle, acceleration slope, leeward vortex index, ridge distance, and valley orientation. To align with the time-series data, this base map was copied 96 times along the time axis, forming a 96×1024×1024×6 topographic dynamic cube, thus obtaining the topographic data.

[0023] S102. Project and overlay the meteorological data, remote sensing data and topographic data respectively to obtain the corresponding data cube.

[0024] In this embodiment, meteorological data, remote sensing data, and topographic data are projected and overlaid to obtain corresponding data cubes. For example, after obtaining wind meteorological data, remote sensing data, and topographic data, all data are uniformly converted to timestamps in the East 8 time zone. The data is then time-series normalized with a fixed target time step of 15 minutes. Simultaneously, complete data from the most recent 24 hours is backtracked, ultimately forming a standardized time-series dataset containing 96 frames, ensuring that all data correspond one-to-one in the time dimension. If the number of consecutive missing frames is less than or equal to 2 (i.e., the missing duration is less than or equal to 30 minutes), linear interpolation is used for filling. If the number of consecutive missing frames is greater than or equal to 2 (i.e., the missing duration is greater than 30 minutes), a convolutional long short-term memory network model is used for forward prediction filling, ultimately ensuring that the entire time-series dataset is continuous and uninterrupted in the time dimension. Then, Albers equal-area conic projection is used to process the meteorological data, remote sensing data, and topographic data respectively, forming a unified spatiotemporal lattice, i.e., the corresponding data cube, which includes a remote sensing cube. Terrain cube and weather cube Where T is the time step, This refers to the number of pixels vertically. This refers to the number of pixels horizontally. For the number of channels, The set is a real number set. At this point, the three data cubes do not overlap channels; only the spatiotemporal dimensions are aligned. The core parameters of the projection are the central meridian 105°E, the standard parallels 25°N and 35°N, with a spatial dimension of a 1024×1024 pixel square and a raster resolution of 1km×1km. Specifically, for the strong wind meteorological data acquired from ground meteorological stations, a deep Gaussian process sparse regression algorithm is used for spatial interpolation to expand the discrete strong wind meteorological data into continuous 1km resolution areal data. Then, cross-validation is used to verify the data accuracy. The results show a root mean square error of 1.8m / s, ensuring that the reliability of the interpolated areal data meets application requirements. For remote sensing data, guided by visible light band data at 0.5km resolution, an enhanced super-resolution generative adversarial network algorithm is used to perform 4x super-resolution reconstruction on other low-resolution bands (such as the infrared channel), uniformly upscaling all bands to 1km resolution. A bilinear reprojection operation is then performed on the super-resolution 1km remote sensing data to ensure its projection coordinate system is consistent with the ground wind and terrain data, ultimately achieving spatial alignment of multi-source data. For terrain data, a cubic convolution interpolation algorithm is used for downscaling, downsampling the original 30m resolution high-precision raster terrain data to 1km resolution while preserving mountain / valley extreme values. After the above preprocessing, a spatiotemporal unified cubic tensor corresponding to meteorological data, remote sensing data, and terrain data is obtained, i.e., a data cube, with the shape of... That is, 96 time steps, with 15-minute intervals, 1024 vertical pixels, 1024 horizontal pixels, and 15 channels. Among them, the 15 channels include 8 satellite brightness temperature channels, 6 terrain dynamic parameter channels, and a station ground truth gust sequence. The data of the above 15 preprocessed channels are used as input data for the two-dimensional embedding layer in the improved visual Mamba model.

[0025] S103. Perform zero-padding and convolution segmentation operations on each data cube to obtain the spatiotemporal image block cube corresponding to each data cube.

[0026] In this embodiment, zero-padding and convolution segmentation operations are performed on each data cube input to the convolutional network. The original large-scale data cube is cut into independent small data blocks according to the convolution kernel size, resulting in corresponding spatiotemporal image block cubes. During convolution, the corresponding convolution kernel parameters can be selected according to the characteristics of the meteorological data cube, the remote sensing data cube, and the terrain data cube. For example, the three-dimensional convolution kernel of the meteorological data cube can be set to K=(2,4,4), and the convolution kernel sliding step S is equal to the kernel size K, i.e., S=K. Therefore, there is no overlapping area between adjacent image blocks, ensuring that the data division is regular and without redundancy. The size of each image block is defined as 30min×4km×4km. Temporally, 30 minutes can cover the key period of gust development and cloud evolution. Spatially, 4km×4km can completely capture the coupling relationship between canyon narrow terrain, cloud shape, and gusts, avoiding the loss of local correlation due to excessive size and the omission of key features due to insufficient size. Level pixel compression to Level-based feature units are used, taking computational efficiency into account. Specifically, firstly, an independent and learnable 3D convolutional kernel is applied to each data cube, with its size and stride set synchronously, as shown in the formula: in, This represents the size vector of the 3D convolution kernel. Indicates the length of the timeline coverage. and These represent the height and width of the spatial axis, respectively. This represents the stride vector of a 3D convolution.

[0027] For example, Taking 2 corresponds to a satellite cloud image evolution window of 2×15min=30min, which can completely cover the initial stage of the canyon squall line to its peak. and All values ​​are set to 4, corresponding to a square receptive field of 4×1km=4km, which matches the average bottom width of a canyon in a certain region, ensuring that a single convolution can capture the narrow tube-acceleration local coupling.

[0028] It should be noted that the improved visual Mamba model uses non-overlapping sliding. This setting indicates that the output spacing is 30 min × 4 km. This setting immediately compresses the output tensor size to 1 / 32 of the data cube's pixel size, thereby reducing the... The original pixel count was reduced to 10 in one go. 4 Level feature unit.

[0029] like Figure 2 As shown, Figure 2This is a technical roadmap for multimodal data fusion methods for extreme winds on transmission lines. Based on the determined convolution kernels and strides, zero-padding and convolution segmentation operations are performed on each data cube to obtain the corresponding spatiotemporal image patch cube, i.e., the remote sensing spatiotemporal image patch cube. Topographic spatiotemporal image block cube and meteorological spatiotemporal image block cube Simultaneously, bias is disabled to reduce the number of parameters and avoid positional shifts; subsequent bias is uniformly added by a linear projection layer. Positional information is independently encoded in later steps to ensure segmentation without contaminating the data source. Specifically, zero-padding is performed on each data cube, and then the dimensions are used... Step length The three-dimensional convolution kernel performs a non-overlapping scan, and its formula is: in, A cube of remote sensing spatiotemporal image blocks. For terrain spatiotemporal image block cube A meteorological spatiotemporal image block cube, This represents a 3D convolution operation that can simultaneously perform sliding weighted summation on the time, height, and width axes. For size, Step size, , and These are the convolution window parameters corresponding to the remote sensing spatiotemporal image patch cube, the terrain spatiotemporal image patch cube, and the meteorological spatiotemporal image patch cube, respectively. For remote sensing cubes, For terrain cubes, It is a weather cube.

[0030] S104. Perform a linear transformation on each spatiotemporal image block cube to obtain the corresponding feature unit sequence.

[0031] In this embodiment, after obtaining the spatiotemporal image cubes of each data point—namely, the meteorological spatiotemporal image patch cube, the remote sensing spatiotemporal image patch cube, and the topographic spatiotemporal image patch cube—each data point's spatiotemporal image cube is passed through a dedicated linear layer to generate three independent feature unit sequences: the remote sensing feature unit sequence, the topographic feature unit sequence, and the meteorological feature unit sequence. All sequences have an N×D shape, where N represents the total number of local image patches, and D is the embedding dimension, with N≈3.1×10⁻⁶. 4D=384. These three independent feature unit sequences are not added to each other; they only map each local image patch to a 384-dimensional feature vector through the coding layer, preserving the pure physical semantics for subsequent network fusion. Specifically, firstly, to reduce the dimensionality of the feature units and introduce a learnable metric, a weight matrix and bias are used, with the following formula: In the formula, For embedded dimensions, The 384-dimensional dimension conforms to the channel width of all subsequent SSM (State Space Model) layers.

[0032] Then, perform the following operation on each image block, specifically: In the formula, , and These are the learnable weight matrices corresponding to the remote sensing spatiotemporal image patch cube, the terrain spatiotemporal image patch cube, and the meteorological spatiotemporal image patch cube, respectively, with dimensions of 480×384. A cube of remote sensing spatiotemporal image blocks. A cube of spatiotemporal topographic image blocks. A meteorological spatiotemporal image block cube, It is a sequence of remote sensing feature units. It is a sequence of terrain feature units. It is a sequence of meteorological characteristic units. , and These are the bias terms corresponding to the remote sensing spatiotemporal image block cube, the topographic spatiotemporal image block cube, and the meteorological spatiotemporal image block cube, respectively.

[0033] Finally, the output sequences of remote sensing feature units, topographic feature units, and meteorological feature units are given, with dimensions of [size missing]. Where D=384, N=48×256×256≈3.1×10 4 .

[0034] S105. Using the network coordinates of each data block in each spatiotemporal image block cube as the spatiotemporal index, add positional encoding to the corresponding feature unit sequence to obtain the corresponding final feature unit sequence. The corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence.

[0035] In this embodiment, since the improved visual mamba model lacks image patch location information by default, a learnable 3D location embedding needs to be added. Specifically, such as... Figure 2 As shown, Figure 2 This is a technical roadmap for multimodal data fusion methods for extreme winds on transmission lines. In each feature unit sequence, each feature unit corresponds to a local image block. The spatiotemporal coordinates of each image block within the spatiotemporal image block cube are used as the spatiotemporal index to obtain a positional code. The positional codes of each image block in the feature unit corresponding to each image block are added to or concatenated (usually added) to obtain a homomorphic feature unit sequence with added positional codes, i.e., the final feature unit sequence. The specific formula is as follows: in, For position encoding, , and Trainable weights corresponding to each feature unit, , , These are the integer coordinates of the image patch within the time, height, and width grid dimensions. This is the final remote sensing feature unit sequence. This is the initial and final terrain feature unit sequence. The final meteorological feature unit sequence is given by N, which represents the total number of local image patches and D, which is the embedding dimension.

[0036] It should be noted that integer coordinates allow the network to know whether the cloud formation is on the leeward side 30 minutes ago or has just crossed the mountain pass.

[0037] In one embodiment, when obtaining the corresponding final feature unit sequence, the method further includes: By projecting terrain dynamic parameters onto the terrain data cubes in each data cube, higher-order terrain dynamic features are obtained. Higher-order terrain dynamic features are added to the corresponding feature unit sequence to obtain the final terrain feature unit sequence.

[0038] In this embodiment, the 6-dimensional terrain dynamic parameters in the data cube obtained in step S01 are... Another 1×1×1 convolution is projected onto the same embedding dimension, as shown in the formula: In the formula, The topographic dynamic tensor is a 6-dimensional tensor, consisting of aspect ratio, shading angle, acceleration slope, leeward vortex index, ridge distance, and valley orientation. These are the weight parameters for 3D convolution. This represents the convolution operation. This indicates the output terrain dynamic feature unit, i.e., the high-order terrain dynamic feature, which is in the same dimension as the terrain feature unit sequence. This represents the total number of local image patches, where D is the embedding dimension. =48×256×256≈3.1×10 4 D=384.

[0039] Finally, only topographic feature unit sequences are injected; meteorological and remote sensing feature unit sequences are not injected. This aims to maintain the purity of cloud image evolution and ground truth anchors. The specific formula is as follows: In the formula, This is the final sequence of terrain feature units. This indicates the output terrain dynamic feature unit. This is the initial and final terrain feature unit sequence.

[0040] Specifically, in the final remote sensing feature unit sequence, the compressed representation of the brightness and temperature channels of eight satellites within a 30min×4km×4km area (384-dimensional vector per row) has been incorporated into the location encoding corresponding to the image patch. The final remote sensing feature unit sequence includes cloud image evolution and spatiotemporal coordinate information. Each row of data in the final topographic feature unit sequence includes six topographic dynamic parameters, topographic-specific location encoding, and an output topographic dynamic feature unit, also known as the leeward vortex residual, which includes canyon sluices and soft constraints on acceleration coefficients. Leeward vortices are the turbulent characteristics of airflow after passing through a canyon; the residual reflects its deviation from the ideal sluice effect. Through this physical prior constraint, the model learns how canyon topography accelerates airflow (gusts) through the sluice effect. Each row of the final meteorological feature unit sequence is 384-dimensional, including 1km gust values, ground truth uncertainty, and station-specific location encoding. The final remote sensing feature unit sequence, final topographic feature unit sequence, and final meteorological feature unit sequence, as the final products of the embedding layer, are input into the subsequent two-branch feature extraction module.

[0041] S106. Based on the obtained final feature unit sequences, perform unique flow extraction and common flow extraction respectively to obtain common flow features and unique flow features.

[0042] In this embodiment, a two-branch feature extraction process—unique flow and shared flow—is performed within the backbone of the improved visual Mamba model. Specifically, physical hard index extraction and cross-source covariance-gated two-branch feature extraction are completed first: the unique flow allows each feature unit sequence to refine 192 irreplaceable hard features before the Mamba layer, including: cloud dynamic evolution features (convective intensity), terrain acceleration (canyon aspect ratio, leeward slope, and acceleration coefficient values), and extreme value jumps (gust jumps greater than 5ms within 15 minutes). -1 (Pulse). A 192-dimensional spatiotemporal skeleton and covariance-guided weights are extracted from the common flow and modulated into a unique flow to facilitate subsequent data fusion processing.

[0043] In one embodiment, based on the obtained final feature unit sequences, unique flow extraction and common flow extraction are performed respectively to obtain common flow features and unique flow features, including: Unique flow extraction is performed on the final remote sensing feature unit sequence to obtain the dynamic evolution characteristics of cloud clusters; Unique flow extraction is performed on the final terrain feature unit sequence to obtain terrain features; The meteorological features are obtained by extracting unique flows from the final meteorological feature unit sequence. The dynamic evolution characteristics, topographic features, and meteorological features of cloud clusters were identified as unique flow characteristics. Shared flow features are extracted based on the final remote sensing feature unit sequence, the final topographic feature unit sequence, and the final meteorological feature unit sequence.

[0044] In one embodiment, unique flow extraction is performed on the final remote sensing feature unit sequence to obtain the dynamic evolution features of the cloud cluster, including: Feature analysis is performed on the final remote sensing feature unit sequence to obtain a new terrain feature unit sequence; The dynamic evolution patterns of cloud clusters are extracted from the sequence of new terrain feature units to obtain the dynamic evolution characteristics of cloud clusters.

[0045] In this embodiment, firstly, before the first layer SSM of the improved visual mamba model, the unique information, such as the brightness temperature decrease rate, is extracted from the final remote sensing feature unit sequence to avoid dilution by terrain or site information. Figure 3 As shown, Figure 3 The diagrams show the extreme value jump self-attention network module and the SE_Block network architecture. The final remote sensing feature unit sequence is processed using the time-dilated SSM and SE_Block networks, with the following formula: in, This represents the time-axis expansion factor, increasing the receptive field from 30 minutes to 90 minutes, capturing the complete lifecycle of the canyon squall line. This indicates the attention of the Squeeze-Excite channel. This refers to the dynamic evolution characteristics of cloud clusters. This describes the dynamic evolution characteristics of cloud clusters.

[0046] Specifically The network's input feature map is the final remote sensing feature unit sequence, with channel C having a dimension of 384, 48 temporal image patches, and a spatial image patch resolution of 256×256. First, global average pooling is performed on each 384-dimensional channel to obtain a 384-dimensional channel description vector. Then, 0-1 384-dimensional channel weights are generated through dimensionality reduction, dimensionality increase fully connected pairs, and a sigmoid function. Finally, the weights are multiplied back into the original feature map channel by channel, achieving lightweight attention that enhances strong convection kernel channels and suppresses redundant channels, while maintaining the output dimension. The shape is obtained, which is the dynamic evolution feature of the cloud cluster. Then, it enters the time-dilated SSM network to continue to extract the cloud evolution feature, and the dynamic evolution feature of the cloud cluster is determined to be one of the unique flow features. The network automatically weights the brightness temperature data of 8 bands, additionally sets the brightness temperature to less than −60°, and sets the channel weight to twice the original weight.

[0047] In one embodiment, unique flow extraction is performed on the final terrain feature unit sequence to obtain terrain features, including: The final sequence of terrain feature units is deconvolved to obtain the first spatial map; The slope component is obtained by performing slope convolution on the first spatial graph in multiple directions using multiple direction kernels. By concatenating the slope component and the Laplace elevation residual, a two-dimensional physical vector of the terrain is obtained. Terrain features are obtained by convolving and projecting two-dimensional physical vectors.

[0048] In this embodiment, before the first layer SSM of the improved Visual Mamba model, the final terrain feature unit sequence first extracts irreplaceable hard indicators such as canyon depth-to-width ratio, leeward slope, and acceleration coefficient. Specifically, four-directional convolution is first used to capture the anisotropy of the airflow over the mountain, then the slope component along the 700hPa wind direction is spliced ​​with the Laplace operator elevation residual to form a two-dimensional physical quantity, and finally a 1×1 convolution projection is used to obtain a 192-dimensional hard feature. This refers to terrain features, thereby enabling the interpretability, verifiability, and moduloability of terrain-accelerated physical quantities. The systematic deviation of leeward wind speed is reduced by 5ms. -1 Reduced to 1.4ms -1The computational cost is less than 0.1%. Specifically, a four-directional slope convolution (physical receptive field) operation is performed. First, the final terrain feature unit sequence is deconvolved back into a 1km spatial map to obtain the first spatial map, ensuring that each pixel can be computed for physical quantities. The deconvolution formula is: In the formula, This is the first spatial map, specifically the 1km elevation sub-map. It is a deconvolution function.

[0049] It should be noted that the 1km elevation submap consists of 48 frames, with 256×256 grid points.

[0050] Secondly, four convolutional kernels with different directions are designed, corresponding to horizontal, oblique, vertical, and anti-oblique directions, respectively. A convolution operation is then performed on the first spatial map to extract terrain slope and undulation features in different directions, thereby capturing the differences in airflow movement in each direction. For example, a 90° convolutional kernel extracts terrain obstruction features in the vertical wind direction, and a 45° convolutional kernel extracts the uphill features of oblique airflow. The convolution formula is as follows: in, The convolution output for each direction, To stack along the channel dimension, It is a two-dimensional convolution function. It is a 3×3 oriented core. Fill with zeros.

[0051] Then, the slope component and the Laplace elevation residual are concatenated into a two-dimensional physical vector to ensure that each pixel has an interpretable physical quantity. The formula is as follows: in, Represents a two-dimensional physical vector of terrain. This represents the slope component along the prevailing wind direction at 700 hPa, in degrees, and is calculated as follows: , The azimuth angle of the wind direction at 700 hPa; ; The second-order Laplace elevation is given by units of 1. , , , These represent the two components of the unit wind direction vector. , .

[0052] It should be noted that the 700 hPa wind azimuth is taken from the hourly atmospheric vertical stratification data with a spatial resolution of 0.25° × 0.25° in the climate reanalysis dataset. The 700 hPa wind azimuth is then bilinearly interpolated to a 1 km Albers grid and then... Calculate the azimuth angle.

[0053] Finally, the 2D terrain physical vector is convolved and projected using a 1×1 method to reduce the dimensionality to 192, as shown in the formula: in, The learnable weight matrix corresponding to the terrain spatiotemporal image patch cube is initialized and then iterated step by step. This indicates no spatial mixing, only a linear channel mapping. These are terrain features used to calculate the terrain magnification factor using unique information provided by the digital elevation model. Represents a two-dimensional physical vector of the terrain.

[0054] In one embodiment, unique flow extraction is performed on the final meteorological feature unit sequence to obtain meteorological features, including: The final meteorological feature unit sequence is deconvolved to obtain the second spatial map; Based on the second spatial map, the meteorological two-dimensional physical vector is obtained; The attention map is obtained by extracting gust jump features based on meteorological two-dimensional physical vectors. The attention map is convolved and projected to obtain meteorological features.

[0055] In this embodiment, before the first layer SSM of the improved Visual Mamba model, the gust rise value and uncertainty within the 15-minute time step of the final meteorological feature unit sequence are concatenated into a two-dimensional physical quantity, wherein the gust rise value ranges from 0 to 15 ms. -1 The uncertainty range is 0-3ms -1 Then, a 192-dimensional vector is projected using 3×3 spatial self-attention and 1×1 convolution. Each number can be deduced to retrieve the specific jump value and uncertainty, serving as an irreplaceable anchor point for the truth value and a hard indicator of extreme jumps. It can be understood that the 192-dimensional vector is not an abstract feature, but a structured encoding of a two-dimensional physical quantity. The true jump value and error range can be restored through inverse operations, ensuring the interpretability of the features. Specifically, the final meteorological feature unit sequence is first deconvolved back into a 1km spatial map, resulting in the second spatial map. This ensures that each pixel can be calculated for gust jumps and uncertainties, using the following formula: in, This is the second space diagram. It is a deconvolution function. For the direction core, The stride of the deconvolution is... It is a sequence of meteorological characteristic units.

[0056] Secondly, construct gust and uncertain two-dimensional physical vectors to ensure that each pixel has an interpretable physical quantity, as shown in the formula: in, This represents the gust value over 1km, in milliseconds (ms). -1 The source is the surface field data in step S101; σ represents the variance of the depth Gaussian process, in milliseconds. -1 If σ is greater than 1.5, the interpolation blind zone is marked. It is a two-dimensional physical vector for meteorology.

[0057] Then, since gust jumps have spatial continuity, an extreme jump self-attention spatial correlation operation is performed, using a 3×3 directional kernel to capture gust jumps greater than 5ms within 15 minutes. -1 The pulse, the formula is: in, This represents a 3×3 directional kernel, with dilation=3 indicating a receptive field of 9km, matching the spatial scale of canyon gusts; A_stat∈R 48×256×256×2 For extreme values, jump to uncertain attention maps. For extreme value jumps, the attention space association operation is performed. It is a two-dimensional physical vector for meteorology.

[0058] Finally, the two-dimensional physical quantities are projected onto the 192-dimensional model space, preserving the magnitude relationships of the physical quantities but removing numerical redundancy between pixels. The formula is: in, represents the learnable weight matrix corresponding to the meteorological spatiotemporal image patch cube; Conv1×1 represents no spatial mixing, only channel linear mapping; This is an extreme value jump characteristic, i.e., a meteorological characteristic. For extreme values, jump to uncertain attention plots.

[0059] In one embodiment, common flow features are extracted based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain common flow features, including: Spatiotemporal skeletons were extracted from the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain coordinate spatiotemporal skeletons. Based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence, cross-source covariance guided weighting operations are performed to obtain gated weights; The common guiding signal is obtained by multiplying the gating weights and the coordinate spatiotemporal skeleton element by element. The shared guiding signal is identified as a shared flow characteristic.

[0060] In this embodiment, before the first layer SSM of the improved Visual Mamba model, the values ​​of the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence are not spliced. Only the spatiotemporal coordinates and the common signal of the three-source covariance are extracted as 192-dimensional guiding weights to modulate the unique flow. This achieves priority fusion of high-value areas simultaneously with abrupt cloud changes, steep terrain changes, and extreme value jumps, avoiding ineffective fusion in plain areas, and obtaining... This refers to the 192-dimensional spatiotemporal skeleton and the shared guiding signal across source covariance. It should be noted that the high-value region simultaneously experiencing cloud abrupt changes, terrain shifts, and extreme value jumps refers to the image patch region where these three phenomena occur simultaneously. The specific process for shared flow extraction is as follows: First, the spatiotemporal skeleton is extracted. A pure spatiotemporal coordinate skeleton is extracted, excluding in-source values ​​to avoid becoming a second concatenation vector. The formula is: in, Represents the integer coordinates of the image patch, t∈[0,47], h,w∈[0,255]; , and For trainable weights, It is the same spatiotemporal skeleton for the three source data, namely the coordinate spatiotemporal skeleton.

[0061] Then, a cross-source covariance guided weight (shared gating) operation is performed. The pixel-level covariance matrices of the feature unit sequences from three different sources are mapped to 192-dimensional gating weights G, retaining only high-value regions with abrupt changes in cloud imagery, steep terrain changes, and extreme value jumps. The calculation process of the gating weights is as follows: in, Represents the pixel-level covariance matrix. This is a function for calculating the pixel-level covariance matrix, computed along the channel dimension. This represents a linear projection from 3D to 192D. The hyperbolic tangent activation function is used. For gating weights, The weights are the pixel-level covariance matrix. This is the final remote sensing feature unit sequence. This is the final meteorological feature unit sequence. This is the final sequence of terrain feature units.

[0062] Finally, the coordinate spatiotemporal skeleton is multiplied element-wise with the gating weights to obtain a common guiding signal for the 192-dimensional spatiotemporal skeleton and covariance, which is used to guide the subsequent fusion of terrain features, meteorological features and cloud evolution features.

[0063] Here, ⊙ represents element-wise multiplication. This is a shared guiding signal for the 192-dimensional spatiotemporal skeleton and covariance. For gating weights, It is the same spatiotemporal skeleton for the three source data, namely the coordinate spatiotemporal skeleton.

[0064] S107. Using common flow features as keys and unique flow features as queries, perform fusion to obtain the fusion result.

[0065] In this embodiment, as Figure 4 As shown, Figure 4 The diagram shows the Cross-Mamba feature fusion network architecture for a multimodal data fusion method for transmission lines under extreme wind conditions. In the improved visual Mamba model, a shared guiding signal is used. Using K as the key and V as the value, and the unique stream as the query Q, a cross-attention module is executed to achieve physically interpretable fusion of shared signal modulation and unique signal modulation. First, define the input data, let Q = [ ; ; ]∈R N×576 ,in, As a topographic feature, Characteristics of cloud evolution, For meteorological characteristics. Then, the query Q, key K, and value V are input into the cross-attention module for cross-fusion, with the formula as follows: in, This represents the bidirectional SSM layer of the improved Visual Mamba model, i.e., the cross-attention module, with K and V as keys and Q as the query. This is the result of the fusion.

[0066] Specifically, the Bi_SSM cross-attention module is embedded between the first and second SSM layers of the improved visual Mamba model: the input layer feeds the unique hard index pool and the shared guiding signal into the linear mapping layer to obtain the query weight matrix W_Q∈R. 576×576 and the key-value weight matrix W_K / V∈R 192×576 Wherein, the query weight matrix W_Q∈R 576×576 and the key-value weight matrix W_K / V∈R192×576 Both are 576-dimensional, ensuring that the unique hard indicator pool and the shared guiding signal are of the same dimension, facilitating subsequent cross-calculation. Then, the improved Visual Mamba model's built-in bidirectional SSM layer is used. The forward SSM uses the shared guiding signal as the key and the unique hard indicator as the query, while the reverse SSM also uses the shared guiding signal as the key and the unique hard indicator pool as the query. The forward and reverse outputs are averaged and the residuals are added back, resulting in a 576-dimensional fusion result. Each element can be linearly deduced to return the specific jump value, terrain coefficient, or uncertainty. Finally, cross-attention is performed pixel-by-pixel and channel-by-channel, using 0-1 weights as the key and the unique hard indicator as the query. When the 0-1 weight is less than 0.5, it indicates that the unique hard indicator is suppressed (plain areas, areas without abrupt changes); when the 0-1 weight is greater than 0.5, it indicates that the unique hard indicator is enhanced, i.e., areas with steep canyons, abrupt cloud changes, and extreme jumps all at high values; when the 0-1 weight equals 1, it indicates that the unique hard indicator is fully enhanced, i.e., areas with steep canyons, abrupt cloud changes, and extreme jumps all at high values. This module uses a unique hard index based on shared covariance weight modulation. It does not splice or mix, thus preserving physical interpretability. Residual connections prevent gradient vanishing and ensure that the hard index is not diluted.

[0067] The multimodal data fusion system for extreme winds on transmission lines provided in this embodiment of the invention, such as... Figure 5 As shown, Figure 5 The system block diagram for the 500 multimodal data fusion system for extreme winds on transmission lines includes: The acquisition module 501 is used to acquire wind meteorological data, remote sensing data and topographic data of the area to be monitored; The projection overlay module 502 is used to project and overlay meteorological data, remote sensing data and terrain data respectively to obtain their respective data cubes. The convolution segmentation module 503 is used to perform zero-padding and convolution segmentation operations on each data cube to obtain the spatiotemporal image block cube corresponding to each data cube. The linear transformation module 504 is used to perform linear transformations on each spatiotemporal image block cube to obtain the corresponding feature unit sequence. The location encoding module 505 is used to add location encoding to the corresponding feature unit sequence using the network coordinates of each data block in each spatiotemporal image block cube as the spatiotemporal index, so as to obtain the corresponding final feature unit sequence. The corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence. The feature extraction module 506 is used to perform unique flow extraction and common flow extraction respectively based on the obtained final feature unit sequences to obtain common flow features and unique flow features; The fusion module 507 is used to fuse common flow features as keys and unique flow features as queries to obtain fusion results.

[0068] The specific implementation of the multimodal data fusion system for extreme winds on transmission lines is basically the same as the specific implementation of the multimodal data fusion method for extreme winds on transmission lines described above, and will not be repeated here.

[0069] In one embodiment of this application, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above steps. The implementation principle and technical effects of the computer device provided in this embodiment are similar to those of the above method embodiments, and will not be repeated here.

[0070] In one embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, it performs the above steps; the implementation principle and technical effects of the computer-readable storage medium provided in this embodiment are similar to those of the above method embodiments, and will not be repeated here.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A multimodal data fusion method for extreme winds on power transmission lines, characterized in that, include: Acquire gale meteorological data, remote sensing data, and topographic data of the area to be monitored; The meteorological data, remote sensing data, and terrain data are projected and superimposed to obtain their respective data cubes; Each of the data cubes is subjected to zero-padding and convolution segmentation operations to obtain the spatiotemporal image block cubes corresponding to each of the data cubes. A linear transformation is performed on each of the spatiotemporal image block cubes to obtain the corresponding feature unit sequence; Using the network coordinates of each data block in each spatiotemporal image block cube as a spatiotemporal index, positional encoding is added to the corresponding feature unit sequence to obtain the corresponding final feature unit sequence, wherein the corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence; Based on the obtained final feature unit sequences, unique flow extraction and common flow extraction are performed respectively to obtain common flow features and unique flow features; The shared flow features are used as keys, and the unique flow features are used as queries to perform fusion, resulting in a fusion result.

2. The multimodal data fusion method for extreme winds on transmission lines as described in claim 1, characterized in that, The process of obtaining the corresponding final feature unit sequence also includes: By projecting terrain dynamic parameters onto the terrain data cubes in each of the data cubes, higher-order terrain dynamic features are obtained. The higher-order terrain dynamic features are added to the corresponding feature unit sequence to obtain the final terrain feature unit sequence.

3. The multimodal data fusion method for extreme winds on transmission lines as described in claim 1, characterized in that, The step involves extracting unique flow and common flow based on the obtained final feature unit sequences, to obtain common flow features and unique flow features, including: Unique flow extraction is performed on the final remote sensing feature unit sequence to obtain the dynamic evolution features of the cloud cluster. Unique flow extraction is performed on the final terrain feature unit sequence to obtain terrain features; Unique flow extraction is performed on the final meteorological feature unit sequence to obtain meteorological features; The dynamic evolution characteristics of the cloud clusters, the topographic features, and the meteorological features were determined to be unique flow characteristics. Common flow features are extracted based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain common flow features.

4. The multimodal data fusion method for extreme winds on transmission lines as described in claim 3, characterized in that, The process of extracting unique flow from the final remote sensing feature unit sequence to obtain cloud dynamic evolution features includes: Feature analysis is performed on the final remote sensing feature unit sequence to obtain a new terrain feature unit sequence; The dynamic evolution pattern of cloud clusters is extracted from the new terrain feature unit sequence to obtain the dynamic evolution characteristics of cloud clusters.

5. The multimodal data fusion method for extreme winds on transmission lines as described in claim 3, characterized in that, The process of extracting unique flow from the final terrain feature unit sequence to obtain terrain features includes: The final sequence of terrain feature units is deconvolved to obtain the first spatial map; The first spatial graph is subjected to slope convolution in multiple directions using multiple directional kernels to obtain multiple convolution results. Based on each of the convolution results, the slope component is obtained. By concatenating the slope component and the Laplace elevation residual, a two-dimensional physical vector of the terrain is obtained. The terrain features are obtained by convolving and projecting the two-dimensional physical vector.

6. The multimodal data fusion method for extreme winds on transmission lines as described in claim 3, characterized in that, The process of extracting unique flow from the final meteorological feature unit sequence to obtain meteorological features includes: The final meteorological feature unit sequence is deconvolved to obtain the second spatial map; Based on the second spatial map, a meteorological two-dimensional physical vector is obtained; Based on the aforementioned meteorological two-dimensional physical vector, gust jump features are extracted to obtain an attention map; The attention map is convolved and projected to obtain meteorological features.

7. The multimodal data fusion method for extreme winds on transmission lines as described in claim 3, characterized in that, The common flow feature extraction based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence yields the common flow features, including: Spatiotemporal skeletons are extracted from the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence to obtain coordinate spatiotemporal skeletons; Based on the final remote sensing feature unit sequence, the final terrain feature unit sequence, and the final meteorological feature unit sequence, a cross-source covariance guided weighting operation is performed to obtain gated weights; The gate weights and the coordinate spatiotemporal skeleton are multiplied element-wise to obtain the common guidance signal; The shared guidance signal is determined to be a shared flow characteristic.

8. A multimodal data fusion system for extreme winds on power transmission lines, characterized in that, include: The acquisition module is used to acquire wind meteorological data, remote sensing data, and topographic data of the area to be monitored; The projection overlay module is used to project and overlay the meteorological data, remote sensing data and terrain data respectively to obtain their respective data cubes; The convolution segmentation module is used to perform zero-padding and convolution segmentation operations on each of the data cubes to obtain the spatiotemporal image block cubes corresponding to each of the data cubes. The linear transformation module is used to perform linear transformations on each of the spatiotemporal image block cubes to obtain the corresponding feature unit sequence. The location encoding module is used to add location encoding to the corresponding feature unit sequence using the network coordinates of each data block in each spatiotemporal image block cube as a spatiotemporal index, so as to obtain the corresponding final feature unit sequence. The corresponding final feature unit sequence includes the final meteorological feature unit sequence, the final remote sensing feature unit sequence, and the final terrain feature unit sequence. The feature extraction module is used to perform unique flow extraction and common flow extraction respectively based on the obtained final feature unit sequences to obtain common flow features and unique flow features; The fusion module is used to fuse the common flow features as keys and the unique flow features as queries to obtain a fusion result.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor is configured to implement the multimodal data fusion method for extreme winds on transmission lines as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the multimodal data fusion method for extreme winds on transmission lines as described in any one of claims 1 to 7.

Images