Wind prediction method, system, device and medium for power grid transmission channel

Abstract

The application relates to the technical field of power system safety protection, and discloses a strong wind prediction method, system, device and medium for a power grid transmission channel, global meteorological features and local topographic features are respectively constructed based on meteorological prediction data and geographic information data of the power grid transmission channel, the obtained two features are input into a composite neural network model constructed based on a convolutional neural network and FiLM modulation for processing, and a wind field change prediction value of the power grid transmission channel under a first spatial resolution is output; based on real-time wind speed data of the power grid transmission channel, an initial moment wind field of the power grid transmission channel under a second spatial resolution is determined by using a double-source distance weighted interpolation method based on terrain correction; a change amount at a prediction moment and a wind value at an initial moment are subjected to vector composition to form a strong wind prediction result; and a km-level spatial resolution, 1-24 hour short-time refined strong wind prediction for a transmission channel in a complex terrain area is realized.

Description

Technical Field

[0001] This invention relates to the field of power system safety protection technology, and in particular to a method, system, equipment and medium for predicting strong winds in power grid transmission channels. Background Technology

[0002] In recent years, as transmission lines have rapidly extended to high altitudes, plateaus, deep mountains, and coastal typhoon areas, wind disasters have shown a significantly increased trend of frequency, suddenness, and localization, further exacerbating the power grid's exposure to natural disaster risks. How to quickly and accurately predict strong winds is a key requirement for improving the disaster defense capabilities of power grid transmission channels.

[0003] Currently, the commonly used wind forecasting scheme in the power grid is the wind speed forecasting method based on time series models, such as the ARIMA model and LSTM network. However, this type of method only uses the time autocorrelation characteristics of the wind field for prediction. It has a simple structure and high computational efficiency, but it cannot reflect the evolution of large-scale weather systems. It has a lag in predicting sudden wind events and is not capable of capturing wind field changes in complex terrain areas. Therefore, it cannot predict strong winds quickly and accurately.

[0004] Therefore, how to quickly and accurately predict strong winds along power grid transmission channels has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for predicting strong winds in power grid transmission channels, solving the problem of how to quickly and accurately predict strong winds in power grid transmission channels.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for predicting strong winds in power grid transmission channels, comprising: Real-time wind speed data, meteorological forecast data, and geographic information data of the power grid transmission channel are acquired, and global meteorological features and local terrain features are constructed based on the meteorological forecast data and the geographic information data, respectively. A composite neural network model is constructed based on convolutional neural network and FiLM modulation. The global meteorological features and the local terrain features are input into the composite neural network model for processing, and the predicted value of wind field change of the power grid transmission channel at the first-level spatial resolution is output. The real-time wind speed data is processed by a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-level spatial resolution. Based on the predicted wind field change and the wind field at the initial moment, the predicted wind conditions for the power grid transmission channel are determined.

[0007] A second aspect of the present invention provides a high wind forecasting system for power grid transmission channels, comprising: The data processing module is used to acquire real-time wind speed data, meteorological forecast data, and geographic information data of the power grid transmission channel, and to construct global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data, respectively. The wind field prediction module is used to construct a composite neural network model based on convolutional neural network and FiLM modulation, and input the global meteorological features and the local terrain features into the composite neural network model for processing, and output the predicted value of wind field change of the power grid transmission channel at the first-level spatial resolution; The wind field construction module is used to process the real-time wind speed data using a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-level spatial resolution. The strong wind prediction module is used to determine the strong wind prediction result of the power grid transmission channel based on the wind field change prediction value and the wind field at the initial time.

[0008] A third aspect of the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the wind prediction method for power grid transmission channels as described above.

[0009] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the wind prediction method for power grid transmission channels as described above.

[0010] Compared with the prior art, the beneficial effects of the embodiments of the present invention are as follows: By using low-resolution global weather forecast data and high-resolution topographic feature data as inputs, this approach not only relies on conventional weather forecast data but also incorporates high-precision geographic information data. This lays the data foundation for fusing global weather and local topographic features, addressing the issues of insufficient consideration for complex terrain and failure to eliminate climatic background interference in existing technologies. A composite neural network integrating FiLM and CNN is used to achieve deep fusion of global upper-air meteorological features and local topographic features, resulting in wind field change trend forecasts with kilometer-level resolution. This solves the problems of large multi-scale information fusion errors and low computational efficiency in existing technologies, ensuring both high-precision prediction and efficient computation. Dual-source interpolation using measured data for topographic correction corrects the bias in wind speed prediction under complex terrain, significantly improving the accuracy of strong wind predictions in complex terrain areas. Attached Figure Description

[0011] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a method for predicting strong winds in power grid transmission channels provided in a certain embodiment of the present invention; Figure 2 This is a schematic diagram of an upper-air meteorological area and a ground target area provided in a certain embodiment of the present invention; Figure 3 This is a structural diagram of a high wind forecasting system for power grid transmission channels provided in a certain embodiment of the present invention; Figure 4 This is a structural diagram of an electronic device provided in a certain embodiment of the present invention; Figure label: Among them, 10 is the data processing module; 20 is the wind field prediction module; 30 is the wind field construction module; 40 is the strong wind prediction module; 5000 is the electronic equipment; 5001 is the processor; 5002 is the bus; 5003 is the memory; and 5004 is the transceiver. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0014] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art will be able to understand the specific meaning of the above terms in this application according to the specific circumstances.

[0015] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is merely for describing specific embodiments and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0016] Short- and medium-term wind forecasts (1-24 hours) are invaluable meteorological elements in power grid operation and maintenance decision-making, but they are also the most prone to errors in forecasting. Strong wind events are characterized by their suddenness, locality, and time sensitivity. Traditional numerical models and products cannot fully meet the refined operational needs of the power grid in terms of spatiotemporal accuracy, often resulting in problems such as delayed warnings, level deviations, and excessively large forecast areas, thus affecting the accuracy of dispatching, maintenance, and operation and maintenance strategies. Improving the accuracy of wind forecasts allows the power grid to take timely measures before strong winds arrive, such as arranging line inspections in advance, adjusting heavily loaded lines, limiting maintenance windows, and strengthening monitoring of key sections, thereby significantly reducing the risk of disasters. At the same time, accurate forecasts can reduce unnecessary power outages or excessive protective measures, lowering operation and maintenance costs and improving the efficiency of human resource utilization. Furthermore, improving short- and medium-term wind speed forecasting capabilities is also an important foundation for achieving refined condition monitoring of power grid equipment, providing more reliable meteorological data for developing resilient power grid construction and extreme weather emergency plans.

[0017] For disaster forecasting, existing technical solutions can be summarized as: "First, forecast wind speed, then select winds of level six or higher from the threshold of level six winds (>10 m / s)." Depending on the differences in wind speed forecasting techniques, existing solutions also include: Wind speed correction methods based on numerical weather prediction (NWP) have become increasingly important as NWP technology has matured, with short-term wind speed forecasts based on models such as WRF, ECMWF, and GRAPES becoming the primary means of prediction. However, NWP still suffers from systematic errors and biases, thus error correction based on numerical forecast results constitutes another important technique. These methods establish a mapping relationship between model output and actual wind speed through statistical modeling or machine learning, correcting NWP wind speeds accordingly. Commonly used methods include linear regression, random forest, XGBoost, and neural networks. The advantage of these methods is that they can utilize the large-scale weather system information contained in NWP, avoiding the lag inherent in time series methods. However, these methods are limited by the spatial resolution of NWP, and their ability to describe the details of local winds in complex terrain remains limited; furthermore, model initialization errors and biases in physical parameterization schemes can lead to prediction instability. Furthermore, when the numerical weather prediction model itself is upgraded, the original correction model may need to be retrained or adjusted, resulting in high maintenance costs; it also still requires computing resources: although the method itself may not require a large amount of historical data, the numerical weather prediction on which it depends requires huge computing resources, and the correction model still needs a continuous stream of real-time numerical weather prediction data as input.

[0018] Hybrid wind speed prediction methods combining time series and non-stationary wind (NWP) data have emerged in recent years to balance the short-term advantages of time series models with the physical consistency of NWP data. For example, the MA–CNN–BiLSTM structure utilizes multi-head attention mechanisms and convolutional neural networks to extract key physical field features such as atmospheric temperature and humidity from NWP data, and then uses a bidirectional LSTM to capture the temporal variation of wind speed, achieving a complementary advantage of both approaches. The advantages of this type of method include: extending the effective extrapolation time of time series models, enhancing the ability to identify sudden wind events, reducing the lag of purely data-driven models, and introducing physical features into the model, which helps improve generalization ability. Although the models in this method exhibit excellent performance, their main drawback lies in their complexity. The model has a complex structure and is difficult to train: Hybrid models integrate multiple heterogeneous components (such as CNN, BiLSTM, and attention mechanisms), resulting in a very complex model structure. This requires careful design and tuning of hyperparameters, which demands a high level of expertise from the developers. The computational cost is high: Whether for model training or real-time prediction, hybrid models require more computational resources and time than single models. There is a risk of overfitting: When the amount of data is insufficient or the quality is low, such a complex model can easily overfit the noise in the training data, thereby causing its generalization performance on unknown data to decline.

[0019] Multi-source wind speed prediction methods based on spatiotemporal correlation are important research directions in recent years because wind speed also exhibits significant autocorrelation in space, especially in complex terrains such as valley winds, slope winds, and accelerated winds in river valleys, where neighboring stations often show similar wind speed change trends. These methods use methods such as Pearson correlation coefficient, mutual information, and grey relational analysis to screen for spatially correlated surrounding stations, incorporating them as auxiliary features into the prediction model to capture the propagation characteristics of regional weather systems and the influence of topography. Multi-source fusion models often employ algorithms such as neural networks, support vector regression, and random forests, significantly improving the accuracy and robustness of wind speed prediction in complex terrain areas. The main drawback of this method stems from its dependence on data quality and spatial relationships. It is highly dependent on station density and data quality: the model's effectiveness is directly related to the density, evenness of distribution, and completeness and accuracy of the surrounding meteorological stations. In areas with sparse stations or frequent data gaps, model performance suffers significantly; spatial relationships are unstable: spatial correlations between stations are not static, but vary with seasons and weather types (such as calm and stable weather versus windy weather). Using fixed correlation coefficients or station combinations may not be suitable for all weather conditions; the ability to capture spatial anomalies is limited: the model mainly relies on historical statistical correlations, and its ability to capture and predict special wind fields that have never occurred historically or that are caused by local minor topography may be insufficient.

[0020] Based on this, in one embodiment, such as Figure 1 As shown, the first aspect of the present invention provides a method for predicting strong winds in power grid transmission channels, comprising: S1. Acquire real-time wind speed data, meteorological forecast data, and geographic information data of the power grid transmission channel, and construct global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data, respectively; wherein, the real-time wind speed data is obtained through dual-source acquisition by tower wind measuring devices and meteorological stations; the real-time wind speed data includes first source wind speed data and second source wind speed data; the geographic information data includes elevation data and land use data; Specifically, this invention collects real-time wind speed data of the power grid transmission channel through a dual-source system—several tower anemometers deployed within the channel and a meteorological station near the channel. The tower anemometers collect the first source of wind speed data, while the meteorological station collects the second. Subsequently, geopotential height and temperature forecast data of the 850 hPa isobaric surface covering a 3-kilometer radius of the power grid transmission channel are extracted from global weather forecast products (such as ECMWF forecast products). The forecast time resolution is typically 1-3 hours, and the spatial horizontal resolution is typically about 1°. This forecast data is used as meteorological prediction data. Simultaneously, a 30-meter resolution digital elevation model (DEM) of the transmission channel is acquired, or geographic information data is extracted from the following data sources: SRTM DEM (Shuttle Radar Topography Mission), with a spatial resolution of 30m or 90m; ASTERGDEM (Advanced Spaceborne Thermal Emission and Reflection Radiometer Global DEM), with a spatial resolution of approximately 30m. Elevation data and land use data (derived from land use type maps, such as forest, farmland, and water bodies) are extracted from these data.

[0021] In one embodiment, constructing global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data respectively includes: The center point coordinates and target area are determined based on the power grid transmission channel, and the upper-air meteorological area is set based on the center point coordinates and the target area; Extract geopotential height and temperature data for a preset time period from the meteorological forecast data; The geopotential height and temperature data are processed separately to obtain geopotential height difference and temperature difference, which are used as the global meteorological features.

[0022] Specifically, firstly, based on the transmission channel's orientation and the distribution of key towers, the geometric center of the channel or the center of the section most frequently affected by strong winds is selected as the "center point" to determine the center point coordinates (Lat, Lon) of the power grid transmission channel, with a spatial resolution of 1 km. Using the center point as a reference, a buffer zone is extended along the transmission channel direction to form a rectangular or polygonal area, ensuring coverage of the entire transmission channel and its surrounding influence range. The resulting target area on the ground is m×n. In this embodiment, the target area is set as a rectangle centered on the center point, spanning 2° east-west (longitude) and 2° north-south (latitude). Based on the target area, considering that the influence of upper-air wind fields on surface winds typically has a large spatial scale, the upper-air meteorological region is set to a slightly larger area than the target area to include more complete large-scale circulation information. In this embodiment, the upper-air meteorological region is set as a rectangle centered on the center point, spanning 5° east-west and 5° north-south (approximately 550 km × 550 km). This range is sufficient to cover the main weather systems affecting the transmission channel (such as upper-level troughs and jet streams), resulting in an upper-air meteorological region of p×q. To ensure sufficient large-scale weather system information is included, the upper-air meteorological region is at least four times the size of the ground target area. A schematic diagram of the upper-air meteorological region and the ground target area is shown below. Figure 2 As shown, it should be noted that Figure 2 This is for illustrative purposes only and does not imply that the grid shown above must be used, nor does it imply that the grid must be square. The forecast start time is denoted as t0, and the forecast target time is denoted as t1.

[0023] The geopotential heights g0 and g1 and temperature data T0 and T1 of the 850 hPa isobaric surface at times t0 and t1 within the upper-air meteorological region p×q are extracted from meteorological forecast data. Spatial difference is performed on the geopotential height field of each pressure layer to calculate the meridional and zonal gradients. The meridional gradient is calculated by dividing the geopotential height difference between adjacent latitudinal grid points at a fixed longitude by the latitudinal distance (converted to a distance) to obtain the geopotential height change per unit distance, reflecting the pressure gradient force in the north-south direction. The zonal gradient is calculated by dividing the geopotential height difference between adjacent longitude grid points at a fixed latitude by the longitude distance (considering cosine correction) to obtain the pressure gradient force in the east-west direction. Finally, the gradients in both directions are combined into a vector field, or used as two independent feature channels, and output as a geopotential height difference feature map. In addition, the magnitude of the gradient (i.e., the total gradient magnitude) can also be calculated to characterize the pressure gradient intensity. Similarly, spatial difference is performed on the temperature field of each pressure layer to obtain meridional and zonal temperature gradients, outputting a "temperature difference feature map." These temperature gradients help identify frontal zones and thermal wind effects. Alternatively, the geopotential height difference Δg = g1 - g0 and the temperature difference ΔT = T1 - T0 can be directly calculated to obtain geopotential height and temperature differences as global meteorological features. These features not only contain the main dynamic and thermal information of upper-level circulation but also highlight gradient information sensitive to wind field changes through difference, providing crucial input for accurately predicting strong winds along power grid transmission channels. Using difference data instead of absolute values ​​effectively eliminates the influence of slowly varying factors such as seasonal changes and background climate fields, allowing the model to focus more on learning the evolution of the weather system itself and enhancing its adaptability to different geographical regions and climate backgrounds.

[0024] This invention effectively captures large-scale circulation dynamics by extracting key physical quantities from the upper-air meteorological field and performing differential processing on them. It reveals the modulation effect of thermal structure on wind field, improves the physical interpretability and computational efficiency of feature expression, enhances adaptability to complex weather processes, and provides more accurate global meteorological features for wind forecasting of power grid transmission channels.

[0025] In one embodiment, the step of constructing global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data respectively further includes: The elevation data is resampled to the target area using bilinear interpolation or cubic convolution to obtain ground elevation data. Based on the ground elevation data, the slopes in the east-west and north-south directions are calculated respectively to obtain the slope angle of the target area; The land use data is spatially resampled using the principle of maximizing area proportion to obtain the cover type of the target area, and the ground roughness is determined based on the cover type. The local terrain features are obtained by combining the ground elevation data, the slope angle, and the ground roughness.

[0026] Specifically, this invention employs bilinear interpolation or cubic convolution to resample the original 30-meter resolution DEM data to a first-level spatial resolution grid (1km × 1km) consistent with the subsequent neural network input. The interpolation method involves: for each 1km grid center point within the target area, using the values ​​of its four nearest 30-meter DEM grid points, calculating the elevation value of that point through distance-weighted averaging, and outputting a "ground elevation data" raster covering the entire target area. Each grid cell contains an elevation value (unit: meters), which is then used as the ground elevation data at the first-level spatial resolution. This resampling process ensures a uniform spatial scale between the terrain features and the model input data, improving the consistency and comparability of subsequent coupled calculations.

[0027] Based on the resampled 1km resolution elevation raster, the central difference method is used to approximately calculate the slope angle (in degrees or radians) of each grid cell in the east-west (longitude) and north-south (latitude) directions, generating two slope rasteres representing the east-west and north-south slope angles respectively. The corresponding slope aspects are obtained and combined to obtain the slope angle dataset for the target area. Furthermore, using the generated ground elevation data, the Horn algorithm, commonly used in Geographic Information Systems (GIS), is employed to calculate the ground slope. The Horn algorithm, based on 3×3 neighborhood elevation differences, calculates the slope in the east-west (dz / dx) and north-south (dz / dy) directions respectively, then calculates the slope angle and determines the corresponding slope aspect. This information is then combined to form the slope angle dataset for the target area.

[0028] Land use data at 30-meter resolution was resampled to a 1-kilometer grid using the principle of maximizing area proportion (i.e., mode resampling). For each 1-kilometer grid, the land use types of all 30-meter pixels within that grid were statistically analyzed, and the type with the largest area proportion was selected as the cover type for that grid. Based on the resampled cover types, a ground roughness length was assigned to each grid according to an empirical parameter table of surface roughness length (referencing relevant meteorological standards or literature), generating a "ground roughness" raster covering the target area, with each grid cell containing a roughness value. Alternatively, the latest global or regional scale land use / land cover (LULC) remote sensing products can be selected as the basic data source, such as MODIS land use and cover products (e.g., MCD12Q1, 500m resolution, annual composite, classification system including IGBP, UMD, and other schemes). Then, spatial resampling was performed using the "maximizing area proportion" principle to determine the land use / cover type of each grid in the target area. Finally, the minimum and maximum ground roughness of various types of cover were read from the vegetation parameter lookup table VEGTLB of the WRF model, and the ground roughness parameter values ​​were calculated using the arithmetic mean method to obtain the ground roughness of the target area.

[0029] The three raster data obtained from the above processing—ground elevation data (1 channel), east-west slope angle (1 channel), north-south slope angle (1 channel), and ground roughness (1 channel)—are stacked in the channel dimension to form a multi-channel local terrain feature tensor. The spatial dimension of this tensor is consistent with the grid division of the target area.

[0030] Furthermore, to facilitate neural network training, the data from each channel is normalized: elevation data can be subtracted from the regional average elevation and then divided by the standard deviation; slope angle can be divided by the maximum possible slope (e.g., 90°) and mapped to the [0,1] interval; roughness can be logarithmically transformed or normalized; and the normalized multi-channel tensor is the "local terrain feature," which will be input together with the global meteorological features into the subsequent composite neural network model for wind prediction of power grid transmission channels. Moreover, for the first-defined forecast area, this invention saves the above three terrain feature parameters as structured data files in GeoTIFF, NetCDF, and HDF5 formats for use in subsequent wind field change forecasts.

[0031] This invention employs bilinear interpolation or cubic convolution to resample the original elevation data to a grid resolution consistent with the target area of ​​the power transmission channel, eliminating the scale difference between the original data and the prediction area, and enabling terrain features to accurately correspond to key points along the line. By calculating the east-west and north-south slope angles, the invention directly reflects the degree and direction of terrain undulation, which is a key factor in correcting local wind field variations. Based on land use data, the invention determines ground roughness, fully considering the weakening effect of different cover types on near-surface wind speeds. By combining elevation, slope, and roughness into a multi-channel input, the invention provides a three-dimensional and comprehensive description of local terrain for the neural network, effectively coordinating it with global meteorological features. This allows the model to grasp both the large-scale circulation background and the modulation effect of local terrain on the wind field, significantly improving prediction accuracy under complex terrain conditions.

[0032] S2. A composite neural network model is constructed based on convolutional neural networks and FiLM modulation. The global meteorological features and the local terrain features are input into the composite neural network model for processing, and the predicted wind field change value of the power grid transmission channel at a first-level spatial resolution is output. The composite neural network model includes a CNN-based global context extractor, a CNN-based terrain feature extractor, a FiLM modulation layer, and an output decoder. In one embodiment, step S2 includes: The global meteorological features are extracted using the global context extractor to obtain a global context vector; The terrain feature extractor is used to extract features from the local terrain features to obtain a terrain feature map; Based on the FiLM modulation layer, the terrain feature map is subjected to dual conditional modulation using the global context vector to obtain modulation fusion features; The output decoder is used to progressively refine the modulation and fusion features, and outputs the predicted wind field change value of the power grid transmission channel at the first-level spatial resolution.

[0033] Specifically, CNN (Convolutional Neural Network) is a supervised learning model. In this invention, CNN is used to extract "global" weather information; FiLM (Linear Modulation of Feature Dimensions) uses the "global" weather (geopotential height and temperature) information extracted by CNN to dynamically and conditionally adjust the behavior of another network (terrain feature data). This invention uses a combination of Convolutional Neural Network (CNN) and Feature Linear Modulation (FiLM) to achieve end-to-end refined wind speed prediction from low-resolution upper-air fields to high-resolution surface wind fields. In summary, in the strong wind event prediction model of this invention, a composite neural network of CNN and FiLM is constructed and trained. Global meteorological features are used as upper-air input data, and CNN is used to extract low-resolution upper-air data features. Local terrain features are used as surface input data, and FiLM is used to inject the low-resolution upper-air data features into the high-resolution terrain features, thereby outputting the predicted wind field change values ​​Δu and Δv of m×n grid resolution for the ground target area, achieving multi-scale information fusion. The technical implementation environment of this invention is a Python environment on a Linux server, and the function libraries called mainly include netCDF4, pandas, PyTorch, NumPy, etc. The model is set to run on the CPU (it can also be switched to a GPU if needed, because in actual environments there may be a lack of GPUs or the GPU model may be outdated).

[0034] The application process of composite neural network models includes: This invention employs a CNN-based global context extractor to extract global weather information from large-scale, low-resolution background weather forecast data—global meteorological features. It compresses the original p×q spatial grid of 850 hPa geopotential height weather forecast data (including geopotential height difference, temperature difference, etc.) into a vector of length `context_dim`. The invention uses four CNN layers for feature extraction, with identical parameters for each layer. The number of input and output channels is two (representing geopotential height and atmospheric temperature, two variables). The CNN layer kernel size is 5×5, and the kernel's stride on the image is set to 2. Image edge padding is also set to 2 to ensure the size is precisely halved after downsampling. After four downsampling passes, the feature map size gradually decreases from p×q to 1×1. Each convolutional layer is followed by a ReLU activation function and a batch normalization layer. Finally, a fully connected layer projects the features onto a global context vector of a specified dimension (`context_dim`).

[0035] This CNN-based terrain feature extractor extracts rich local terrain features from small-scale, high-resolution terrain data—specifically, local terrain features. It uses three convolutional layers for feature processing, each with identical parameters, a 1×1 kernel size, and no image edge padding. The first layer converts the input channels to 4, the second to 6, and the third to the target feature dimension (feature_dim, which can be modified). Each convolutional layer is followed by a Tanh activation function and a batch normalization layer. Finally, an additional residual connection-style convolutional layer is added to enhance feature representation and obtain the terrain feature map.

[0036] The FiLM modulation layer is used to conditionally modulate local terrain features using a global context vector. The modulation formula is: Output = Input × (1 + γ) + β, where γ = Tanh(W γ × global context vector) and β=Tanh(W β ×Global Context Vector), generated from the global context vector through a fully connected layer. This scheme employs a dual modulation mechanism: the first layer performs coarse-grained (low spatial resolution) modulation, and the second layer performs fine-grained (high spatial resolution) fine-tuning, enhancing the model's expressive power and thus obtaining modulated fusion features. During the modulation process, scaling and offsetting operations are performed on the original features to achieve the fusion of global information and local features. Tanh activation function: Tanh(x) = (e x -e -x ) / (e x +e -x This is used in the feature extraction and modulation process to ensure that the function is differentiable everywhere, while restricting the feature values ​​to the range of [-1,1] to prevent the model from not converging and gradient explosion.

[0037] The output decoder is used to generate the final wind field component predictions from the modulated feature map. The decoder employs a progressive feature fusion strategy, gradually refining and compressing the feature dimensions through three cascaded 1x1 convolutional layers. Each layer uses the Tanh activation function to perform a non-linear transformation on the features and constrain their range to adapt to the continuous variation characteristics of the wind field data. The final layer outputs directly without applying an activation function, preserving the positive and negative direction information of the wind field (u, v components). This scheme uses a pure convolutional structure, maintaining a 1km spatial resolution throughout, and ultimately outputs the predicted wind field changes from time t0 to t1.

[0038] This invention employs independent global context extractors and terrain feature extractors to process input data with different physical properties. The global branch focuses on capturing the dynamic and thermal structure of large-scale circulation systems, while the terrain branch specifically learns the static constraints of local terrain. This decoupled design avoids feature interference caused by direct mixing of heterogeneous data, allowing each branch to extract deep features from its respective domain more purely. FiLM modulation (affine transformation) of the terrain feature map is performed using the global context vector, essentially using global meteorological information as a "condition" to dynamically adjust the expression of terrain features. This enables the model to flexibly adjust the degree of terrain influence according to different meteorological backgrounds, significantly improving its modeling ability for complex nonlinear interactions (such as the coupling of terrain forcing and weather systems). The output decoder uses progressive refinement to gradually restore the modulated and fused features to a first-level spatial resolution, effectively fusing multi-scale information and avoiding the jagged effect caused by direct upsampling. This generates a smooth and physically consistent wind field change field, which is particularly suitable for capturing wind speed gradient changes along power transmission channels.

[0039] Furthermore, this invention introduces an "implicit" fusion of physical mechanism constraints into the composite neural network model, primarily geostrophic wind constraints and mass conservation constraints. The geostrophic wind constraint applies physical constraints to the predicted wind field based on the geostrophic equilibrium principle; the theoretical value of geostrophic wind is obtained by calculating the spatial gradient of the geopotential height field; the north-south gradient of the geopotential height in the east-west direction yields the u component, and the east-west gradient of the geopotential height in the north-south direction yields the v component; finally, the predicted wind field is compared with the theoretical value of geostrophic wind to ensure that the prediction results conform to atmospheric dynamics. The mathematical formula is as follows: In the formula, u g v g denoted by , where is the east-west and north-south component of the geostrophic wind (m / s); f is the Coriolis parameter (s). - ¹), f=2Ωsinω, where Ω is the Earth's rotational angular velocity, ω is the latitude; φ is the geopotential height (m² / s²); , This represents the gradient of the potential elevation in the east-west and north-south directions.

[0040] The mass conservation constraint imposes physical constraints on the predicted wind field based on the principle of mass conservation. The horizontal divergence field is obtained by calculating the east-west gradient of wind field u and the north-south gradient of wind field v; according to the mass conservation requirement, the horizontal divergence should be close to zero. This constraint ensures that the predicted wind field maintains reasonable continuity in spatial distribution, conforming to basic principles of fluid mechanics. The mathematical formula is as follows: In the formula, For horizontal divergence; , These represent the eastward and northward components of the wind field.

[0041] In composite neural networks, the "implicit" fusion of physical constraints is achieved during the model training phase. Specifically, during model training, two physical constraint modules—geostrophic wind constraints and mass conservation constraints—are added to the loss function of the composite neural network to guide internal parameter updates. A backpropagation mechanism is also introduced to ensure that the model "learns" the physical constraints during training, thereby implicitly fusing physical constraints into the model.

[0042] The model training process is essentially the process of updating the internal parameters of the composite neural network model. The internal parameters of the neural network model are determined through forward propagation of the model results and backpropagation to update the model parameters. These two steps are briefly described here: Forward propagation: The model extracts features from the training data through a feature extractor, then modulates the features through a FiLM layer, outputs the predicted value, and enters the loss function.

[0043] Backpropagation: Based on the wind speed difference prediction and the actual value, the loss between the forward propagation prediction and the actual value is calculated using the total loss function (value = data loss + physical constraint loss). Then, based on the calculated loss, the model uses backpropagation to find the parameter gradient, and updates the model parameters in the gradient direction according to the set learning rate. To accelerate the convergence speed of the model parameters, an optimization algorithm is used during backpropagation. This algorithm introduces first-order and second-order momentum in the gradient descent direction to find the fastest parameter convergence direction, allowing the model to continuously approach the actual value.

[0044] After updating the parameters through backpropagation, the model calculates the result through forward propagation, then calculates the loss function based on the result, and finally updates the parameters using backpropagation. After multiple iterations, the loss value remains stable within the region, indicating that the model parameter optimization has converged, and the model has completed the training phase.

[0045] Regarding the selection of optimizer, this invention employs the Adam optimizer. Training a neural network is essentially a complex optimization problem: we need to find a set of parameters (weights and biases) that minimizes the loss function. Due to the complexity of the model, we cannot solve it directly; we can only approximate the optimal solution step by step through iteration. The optimizer is the rule that guides this "step-by-step approximation" process. It manifests as an algorithm that guides the updating of neural network parameters to minimize the loss function. Its core task is to determine "how" and "in which direction" to update the weights so that the model can learn more quickly and stably.

[0046] The Adam optimizer in this scheme employs an adaptive learning rate, maintaining an independent and adaptive "effective learning rate" for each parameter in the neural network. This effective learning rate is obtained by multiplying the global learning rate α by a factor that is dynamically adjusted based on the gradient history (magnitude and direction) of each parameter. This dynamically adjusted adaptive learning rate has the following advantages, making it suitable for this invention: it is suitable for handling the sparse gradient problem that may exist in meteorological data; it has a fast convergence speed and is very effective for complex physically constrained neural networks; the adaptive learning rate also has hyperparameter robustness, and the default parameters usually work well, making it suitable for scientific research prototype development; the adaptive learning rate method has been thoroughly tested and verified in the deep learning community, demonstrating good stability.

[0047] This invention also provides a verified scheme for setting the model's extrinsic parameters and the optimal configuration scheme. This scheme pre-sets nine extrinsic parameters for the CNN model, as detailed in Table 1 below. Based on the pre-set schemes for each parameter, various models can be built through permutations and combinations. These models are then trained one by one to calibrate their internal parameters, and the accuracy of each model can be verified using test set data. Furthermore, based on the performance of each model, the extrinsic parameter configuration scheme with the highest accuracy can be selected.

[0048] Table 1 Model External Parameter Setting Scheme and Optimal Configuration Scheme Training Epochs: One epoch represents a complete pass through the neural network of the entire training dataset, ensuring the model fully learns the data features. In this invention (Epoch=8000): Epochs 1-100 learn the basic temperature-wind field relationship; Epochs 101-1000 learn the modulating effect of terrain on the wind field; Epochs 1000+ fine-tune the balance of physical constraints. Batch Processing: A batch is a subset of samples processed simultaneously in each weight update step, providing stable gradient estimation. In this invention (Batch_size=7200): More accurate gradient estimation: The average gradient of 7200 samples has less noise and more accurate direction than the gradient of smaller batches. Smoother convergence path: Reduces the randomness of individual batches, making the training process more stable. Suitable for physically constrained models: Physical laws require a large amount of data to learn accurately. Batch determines the quality of each update, and Epoch determines the sufficiency of overall learning; both jointly determine the final performance of the model.

[0049] Compared to existing AI wind field forecasting models that either do not consider physical constraints or only explicitly consider them, this invention is a "dedicated" model for refined prediction of surface wind fields in complex terrain areas. The model fully considers the environmental factors and mechanisms of change affecting kilometer-level wind fields in complex terrain areas, implicitly embedding physical constraints, thus reducing the model's size and computational load, requiring less storage and computing resources. It innovatively incorporates fundamental physical principles of atmospheric dynamics, embedding prior knowledge into the neural network through geostrophic wind constraints and mass conservation constraints. This ensures that the predicted wind field and geopotential height field in the free atmosphere satisfy the geostrophic equilibrium relationship, and forces the horizontal wind field divergence to be close to zero, conforming to the continuity equations of incompressible fluids. This physics-guided machine learning method not only improves the physical rationality of the prediction but, more importantly, significantly enhances the model's generalization ability in extrapolation scenarios such as extreme weather and sparse data regions. Compared to purely data-driven AI models, physical constraints act as regularization, preventing the model from learning non-physical spurious correlations, ensuring that the prediction results maintain basic atmospheric dynamic characteristics even under unfamiliar weather conditions, resulting in higher reliability. Compared to numerical models, it maintains the rapid reasoning advantage of AI, avoids complex iterative calculations and parameterization schemes, and makes the simulation results interpretable based on displayed data.

[0050] In one embodiment, the step of performing dual conditional modulation on the terrain feature map based on the FiLM modulation layer using the global context vector to obtain modulation fusion features includes: Based on the target region, a coarse adjustment parameter set is determined, and the global context vector is input to the first set of independent fully connected layers. After Tanh activation, the first scaling factor and the first offset factor under the coarse adjustment parameter set are generated. Upsample the first scaling factor and the first offset factor respectively to obtain a sampling scaling factor and a sampling offset factor with the same spatial resolution as the terrain feature map; The terrain feature map is scaled and offset channel by channel using the sampling scaling factor and the sampling offset factor to obtain a coarse-grained terrain feature map with global coarse adjustment. The global context vector is input into a second set of independent fully connected layers. After Tanh activation, a second scaling factor and a second offset factor with the same spatial resolution as the terrain feature map are generated. The coarse-grained terrain feature map is subjected to secondary channel-by-channel scaling and offset processing using the second scaling factor and the second sampling offset factor to obtain locally fine-tuned modulation and fusion features.

[0051] Specifically, this invention employs a dual modulation mechanism, namely a two-level feature modulation strategy of coarse-grained (low spatial resolution) modulation + fine-grained (high spatial resolution) fine-tuning. Relying on the global context vector, it generates independent scaling / offset parameters for terrain features in two steps, achieving feature fusion by first controlling the large-scale patterns of global weather and then fine-tuning the local differences in complex terrain. The entire process is completed based on tensor operations, without information loss caused by spatial interpolation, and all parameters participate in the backpropagation training and optimization of the model. This not only achieves deep integration of global information and local features, but also avoids the accuracy loss caused by spatial interpolation in traditional downscaling methods.

[0052] In coarse-grained modulation, the resolution is much lower than 1km of the terrain feature map. If the terrain feature map is an M×N grid, then the coarse-grained parameter set is set to an M / 4×N / 4 or M / 8×N / 8 grid. The scheme does not specify a specific scaling ratio. The global context vector is input into the first set of independent fully connected layers (W). γ1 W β1 After Tanh activation, γ1 (first scaling factor) and β1 (first offset factor) are generated under a low spatial resolution coarse-tuning parameter set. These two parameters represent the global modulation intensity of various terrain feature channels based on the current meteorological background. The low-resolution γ1 and β1 are upsampled using bilinear interpolation / nearest neighbor interpolation and stretched to the same 1km spatial resolution and tensor dimension as the terrain feature map, forming the sampling scaling factor and sampling offset factor. This ensures that the parameter set and terrain features can be used for tensor operations on a grid-by-grid basis. The sampling scaling factor and sampling offset factor obtained after upsampling are substituted into the FiLM core formula to perform channel-by-channel and grid-by-grid scaling and offset operations on the original terrain feature map, resulting in a coarse-grained terrain feature map after global coarse-grained modulation. This layer only realizes the overall calibration of terrain features by global weather, allowing the large-scale global weather system to perform overall coarse-tuning of terrain features in a holistic manner, determining the global trend of terrain feature modulation, and avoiding subsequent fine-tuning from deviating from the weather background and failing to capture the local modulation differences of small-scale terrains such as canyons, mountains, and slopes.

[0053] To achieve accurate matching of local differences in complex terrain during fine-grained tuning, the same global context vector is input into a second set of independent fully connected layers (W). γ2 W β2(Completely independent of the parameters of the fully connected layer of the coarse-tuned layer, avoiding modulation accuracy loss caused by parameter sharing), after Tanh activation, it directly generates a set of γ2 (second scaling factor) and β2 (second offset factor) with the same 1km spatial resolution and tensor dimension as the terrain feature map, without the need for further upsampling, achieving precise modulation "grid-by-grid, channel-by-channel". Substituting γ2 and β2 into the FiLM core formula, a second channel-by-channel, grid-by-grid scaling + offset operation is performed on the coarse-grained terrain feature map to obtain the final FiLM modulation fusion feature map, which is the modulation fusion feature after local fine-tuning; the γ2 and β2 of this layer can adapt to the local features of complex terrain, allowing global weather information to achieve differentiated modulation on different terrain grids (e.g., the scaling factor of the canyon grid is larger to match the wind field acceleration effect). That is, based on the coarse-tuned terrain features, it achieves localized precise calibration of global weather information on complex terrain. By capturing the weather system modulation differences corresponding to small-scale terrain (e.g., canyon acceleration, mountain shading, valley convergence, etc.), the fusion feature is made more consistent with the local wind field formation of complex terrain.

[0054] Furthermore, the weights of the fully connected layers in the coarse and fine tuning layers, as well as the generated scaling and offset factors, are all trained parameters of the composite neural network. They are iteratively optimized along with the model through backpropagation and the Adam optimizer, allowing factor generation to adapt to the matching relationship between global weather information and local terrain features. Moreover, from the original terrain features to the final fused features, a 1km-level resolution is maintained throughout the process, without downsampling / downscaling operations on feature maps, ensuring the fine scale of subsequent wind field forecasts. In addition, the terrain feature maps and γ / β parameter sets are all uniform four-dimensional tensors (batch dimension × channel dimension × height grid number × width grid number), and all modulation operations are native PyTorch tensor operations, ensuring computational efficiency and model compatibility.

[0055] Traditional downscaling methods typically require interpolation of low-resolution data, a process that introduces additional errors and obscures the original information. This invention employs Feature-wise Linear Modulation (FiLM) technology to achieve intelligent feature modulation from large-scale low-resolution sky data to small-scale high-resolution terrain data. The FiLM layer generates scaling and offset parameters for each feature channel through a global context vector, adaptively calibrating terrain features using a combination of multiplication and addition. This fusion method completely avoids the accuracy loss caused by spatial interpolation while preserving the spatial characteristics of both high- and low-resolution data. Sky information acts as a "control command" guiding the semantic transformation of terrain features, rather than a simple spatial overlay, achieving deep integration of multi-scale meteorological information at the feature level. Compared to single-layer FiLM modulation, the coarse-grained + fine-grained dual modulation of this invention achieves hierarchical feature fusion with global control and local fine-tuning: first, coarse tuning makes the terrain features conform to the evolution law of large-scale weather systems, and then fine tuning captures the local differences of complex terrain. This avoids local modulation from deviating from the weather background and solves the problem that single-layer modulation cannot adapt to the heterogeneity of wind fields in complex terrain. Finally, it achieves lossless deep fusion of low-resolution upper-air weather information and high-resolution local terrain features at the feature level.

[0056] S3. The real-time wind speed data is processed by the terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at the second-level spatial resolution. In one embodiment, step S3 includes: The target region is divided based on the second-level spatial resolution to obtain several target interpolation grids, and the target points in each target interpolation grid are determined. Calculate the horizontal spatial distance from each of the target points to any of the tower wind measuring devices or the meteorological station, and determine the distance weight of each of the horizontal spatial distances using the Gaussian attenuation rule; Based on the potential height, the ground elevation data and the slope angle, the terrain feature differences from each of the target points to any of the tower wind measuring devices or the meteorological station are calculated to obtain the comprehensive terrain similarity weight of each of the target points; Based on the distance weights and the comprehensive terrain similarity weights, the interpolation weights of each target point are determined to perform weighted interpolation processing on the first source wind speed data and the second source wind speed data to obtain the preliminary wind field values ​​of each target point at the second-level spatial resolution. The initial wind field values ​​are combined after being processed by terrain modulation and smoothing to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution.

[0057] Specifically, this invention uses measured wind speed data from power transmission channels as a basis, and integrates two types of measured data—wind measurement devices on power transmission towers and meteorological stations along the line—through a terrain-corrected dual-source distance-weighted interpolation method. At the same time, it incorporates the acceleration and shading effects of complex terrain for fine correction, and then performs smoothing processing along the power transmission channel. Finally, it generates the wind field (east-west u0 and north-south v0 components) at time t0 with a second-level spatial resolution of 50 meters, providing a high-precision benchmark wind field that fits the actual terrain and is based on actual measurements for subsequent superposition of wind field temporal variations (Δu, Δv).

[0058] Taking the power grid transmission channel as the core, the interpolation area (which can also be considered as the target area) covering the line and a certain range along the line is delineated. A spatial grid of 50m×50m is set up to obtain several target interpolation grids. The center point in each grid is the target point P0(x0,y0,z0) for wind field construction.

[0059] For each target point P0, calculate its horizontal spatial distance (Euclidean distance) to any wind measurement point Pi on a tower and a weather station Pj. Calculate the distance weight using the Gaussian attenuation formula. This process is expressed by the following formula: In the formula, For distance weights; d ab σ represents the horizontal distance; σ is the attenuation radius, taken as 2000m for pipeline points and 5000m for meteorological stations.

[0060] Similarly, for each target point P0, the differences in terrain features between it and any wind measurement point Pi on a tower or meteorological station Pj are calculated. The three core terrain factors of relative height, normalized slope, and aspect are integrated, and the similarity weights of each factor are calculated separately and then multiplied to obtain the comprehensive terrain similarity weight. Each factor is calculated using Gaussian decay, and a scaling parameter Lz / Ls / LA (a hyperparameter, the larger the value, the less sensitive it is to differences in terrain features) is introduced. The final overall terrain similarity weight is the product of the weights of the three individual factors. This process is expressed by the following formula: In the formula, z is the relative height after rectangular coordinate conversion (first, a spatial rectangular coordinate system is established with the Earth's center of mass as the origin, and the latitude, longitude, and absolute altitude of the target point / measured point are converted into rectangular coordinates; then, with the target point as the center, a 3×3 50-meter grid neighborhood is selected, which needs to be adapted to the local scale of complex terrain, and the average value of the rectangular coordinate Z value of all grid points in the neighborhood is calculated, and the relative height of the target point is the difference between its own Z value and this average value); S is the normalized slope (first, all 50-meter grid points in the entire transmission channel prediction area are traversed, and the global maximum of the original slope angle is extracted). The values ​​and global minimum values ​​are calculated, and then the original slope angle of each target point is substituted into the minimum-maximum normalization formula to obtain the normalized slope. A is the slope aspect (first calculate the elevation difference of the target point in the 3×3 neighborhood, then substitute it into the slope aspect calculation formula to obtain the original slope aspect angle, convert the original slope aspect angle in radians to degrees, and define it according to the GIS standard: due north is 0°, and the angle increases clockwise, 0°=due north, 90°=due east, 180°=due south, 270°=due west), and the final angle value is the slope aspect). a corresponds to the target point, and b corresponds to Pi or Pj.

[0061] Three key topographic factors affecting wind speed can also be selected: geopotential height (reflecting the background pressure field, which can be obtained from meteorological forecast data at the target point's location, simplified here to be replaced by surface pressure), surface elevation (from the resampled DEM), and slope angle (from the previously calculated east-west and north-south slopes, which can be combined into a total slope). Each factor is then normalized to eliminate the influence of dimensions. For example, the elevation data is subtracted from the regional average elevation and then divided by the standard deviation to obtain the standardized value. For the target point and the observation point, the differences between them on each factor are calculated. The comprehensive topographic differences are calculated using weighted Euclidean distance, and the topographic differences are converted into similarity weights using a Gaussian form.

[0062] Multiplying the distance weights mentioned above by the terrain similarity weights yields the final interpolation weights from each target point to its corresponding data source. These weights directly reflect the contribution of the measured data from that data source to the wind field at the target point. After normalizing the calculated final interpolation weights, a weighted average is calculated for the measured wind speed components (u, v) from both sources. This yields the components u0 and v0 in the preliminary wind field value at time t0 for each 50-meter grid target point, achieving the fusion of the two types of measured data. This process is expressed by the following formula: In the formula, W 1,i The final interpolation weights for the wind measurement device from the target point to the i-th transmission tower; W 2,j The final interpolation weight from the target point to the j-th weather station; u1,i , v 1,i The u (east-west) and v (north-south) wind speed components measured by the wind measuring device on the i-th transmission tower (from the first source wind speed data); u 2,j , v 2,j The u (east-west) and v (north-south) wind speed components measured at the j-th meteorological station (from the second source wind speed data); N 1. N 2 represents the total number of wind measurement devices and meteorological stations on power transmission towers.

[0063] Finally, the preliminary wind field values ​​obtained for each target point are subjected to power-based vertical profile correction (the wind shear index used in this process is determined based on the surface roughness, typically 0.1-0.3) and spatial low-pass filtering smoothing, and then combined according to grid position to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution. This invention uses a terrain-corrected dual-source distance-weighted interpolation method to construct the initial wind field, comprehensively considering distance attenuation and terrain similarity. Through terrain modulation and smoothing, the spatial continuity and terrain rationality of the initial wind field are ensured, providing a reliable benchmark for wind field change prediction.

[0064] In one embodiment, the step of combining the preliminary wind field values ​​after terrain modulation and smoothing to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution includes: Based on the slope angle, the slope data of each target point is determined. Combined with the terrain position index, which reflects the relative elevation of the target point, the east-west component of each preliminary wind field value is corrected to obtain the east-west corrected wind field of each target point. Based on the slope angle, the slope aspect data of each target point is determined. Combined with the actual wind direction angle in the real-time wind speed data, the north-south component in each preliminary wind field value is corrected to obtain the north-south corrected wind field of each target point. By combining the east-west and north-south corrected wind fields, the initial wind field of the power grid transmission channel at a second-order spatial resolution is obtained.

[0065] Specifically, the wind field obtained by preliminary interpolation does not take into account the acceleration effect of complex terrain on the wind field (such as canyons, steep slopes, and windward slopes) and the shading effect (such as leeward slopes and low-lying areas). Therefore, it is necessary to make targeted adjustments to the preliminary values ​​based on the terrain characteristics of the target point.

[0066] This invention determines the slope data of each target point based on the slope angle calculated above. It also collects terrain position indices that reflect the relative elevation of the target point. (For power transmission channels with complex terrain, a 5×5 grid with a 50-meter neighborhood window is selected. Centered on the target point, the average absolute elevation of all grid points within the 5×5 neighborhood, excluding the target point, is calculated. The difference between the absolute elevation of the target point and the average elevation of the neighborhood is used as the terrain location index. If the terrain location index > 0, it indicates that the target point is a local highland, such as a mountaintop or ridge, which is prone to wind acceleration; if the terrain location index < 0, it indicates that the target point is a local lowland, such as a valley or depression, which is prone to wind shading; if it is close to 0, it indicates that the target point is a flat slope with no obvious terrain elevation difference.) These two factors are introduced to correct the east-west component of the initial wind field value of each target point, resulting in the corrected east-west wind field. This process is expressed by the following formula: In the formula, For east-west oriented corrected wind field; , These are empirical coefficients, all of which are 0.1.

[0067] Determine the slope aspect data for each target point based on the slope angle. Combined with the actual wind direction angle in real-time wind speed data (Substitute the u and v components from the real-time wind speed data into the four-quadrant arctangent formula to calculate the initial wind direction angle. Convert the initial wind direction angle in radians to degrees to obtain the meteorological standard wind direction angle.) The north-south component of the initial wind field values ​​at each target point is corrected to obtain the north-south corrected wind field. This process is expressed by the following formula: In the formula, This is a north-south oriented wind field correction; The empirical coefficient is 0.05.

[0068] The wind field value after terrain correction may exhibit local abrupt changes due to grid discrepancies. However, the wind field of the transmission channel has spatial continuity. Therefore, it is necessary to perform a moving average smoothing process on all 50-meter target points along the transmission channel to eliminate local abrupt changes and ensure the physical rationality of the wind field. The smoothing rule is to number the target points along the channel according to the line direction k. Each target point only refers to the corrected wind field value of the previous point (k-1), itself (k), and the next point (k+1), and set a fixed weight: 0.5 for itself and 0.25 for each of the adjacent points. After weighted averaging, the final wind field value is obtained. Then, the two smoothed components of each target point are combined to obtain the initial wind field of each target point at the second-level spatial resolution. Combining these components yields the initial wind field of the power grid transmission channel at the second-level spatial resolution.

[0069] This invention significantly improves the spatial resolution and physical realism of the initial wind field by refining it based on terrain features: The wind field is decomposed into east-west and north-south components, with slope and aspect factors introduced for correction, enabling targeted characterization of different mechanisms by which terrain affects airflow; a terrain position index reflecting the relative elevation of the terrain is introduced, allowing the model to distinguish wind speed variations in different landform units; combined with the actual wind direction angle in real-time wind speed data, the windward or leeward state of the airflow relative to the slope can be determined, thus applying different correction coefficients to the north-south component, effectively simulating the bypass and blocking effects of airflow over mountains, avoiding blind adjustments based solely on terrain factors; the corrected east-west and north-south components are recombined to obtain a complete two-dimensional wind field vector, preserving both wind speed magnitude and accurately reflecting wind direction changes, ensuring the physical continuity and interpretability of the wind field, and providing a high-quality initial field for subsequent high wind prediction.

[0070] S4. Based on the predicted wind field change value and the wind field at the initial time, determine the wind prediction result for the power grid transmission channel. Specifically, perform vector synthesis of the obtained predicted wind field change value and the wind field at the initial time to obtain the wind speed prediction result of the power grid transmission channel at a resolution of 1km at time t1. Finally, compare it with the wind judgment threshold (10m / s). If it exceeds the threshold, it is judged as a strong wind. Output the wind speed prediction result and the wind judgment result as the wind prediction result for the power grid transmission channel.

[0071] This application proposes a method for predicting strong winds in power grid transmission channels, addressing the challenge of rapid and accurate prediction. The method uses low-resolution global meteorological forecast data and high-resolution terrain feature data as inputs. It not only relies on conventional meteorological forecast data but also incorporates high-precision geographic information data, laying a data foundation for fusing global weather and local terrain features. This solves the problems of insufficient consideration of complex terrain and failure to eliminate climatic background interference in existing technologies. By fusing a composite neural network of FiLM and CNN, deep fusion of global upper-air meteorological features and local terrain features is achieved, resulting in a wind field change trend forecast with kilometer-level resolution. This solves the problems of large multi-scale information fusion errors and low computational efficiency in existing technologies, ensuring both high-precision prediction and efficient computation. Furthermore, dual-source interpolation using measured data for terrain correction corrects the bias in wind speed prediction under complex terrain, significantly improving the accuracy of strong wind prediction in complex terrain areas.

[0072] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0073] In another embodiment, such as Figure 3 As shown, a second aspect of the present invention provides a high wind forecasting system for power grid transmission channels, comprising: The data processing module 10 is used to acquire real-time wind speed data, meteorological forecast data and geographic information data of the power grid transmission channel, and to construct global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data, respectively. The wind field prediction module 20 is used to construct a composite neural network model based on convolutional neural network and FiLM modulation, and input the global meteorological features and the local terrain features into the composite neural network model for processing, and output the predicted value of wind field change of the power grid transmission channel at the first-level spatial resolution; The wind field construction module 30 is used to process the real-time wind speed data using a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-level spatial resolution. The strong wind prediction module 40 is used to determine the strong wind prediction result of the power grid transmission channel based on the wind field change prediction value and the wind field at the initial time.

[0074] It should be noted that the various modules in the aforementioned wind forecasting system for power grid transmission channels can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module. For specific limitations regarding the wind forecasting system for power grid transmission channels, please refer to the limitations regarding the wind forecasting method for power grid transmission channels described above; both have the same function and role, and will not be repeated here.

[0075] A third aspect of the present invention provides an electronic device comprising: Processor, memory, and bus; The bus is used to connect the processor and the memory; The memory is used to store operation instructions; The processor is configured to execute operations corresponding to a wind prediction method for power grid transmission channels as shown in the first aspect of this application by invoking the operation instructions.

[0076] In one alternative embodiment, an electronic device is provided, such as Figure 4 As shown, Figure 4 The illustrated electronic device 5000 includes a processor 5001 and a memory 5003. The processor 5001 and the memory 5003 are connected, for example, via a bus 5002. Optionally, the electronic device 5000 may also include a transceiver 5004. It should be noted that in practical applications, the transceiver 5004 is not limited to one type, and the structure of this electronic device 5000 does not constitute a limitation on the embodiments of this application.

[0077] Processor 5001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 5001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0078] Bus 5002 may include a path for transmitting information between the aforementioned components. Bus 5002 may be a PCI bus or an EISA bus, etc. Bus 5002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0079] The memory 5003 may be a ROM or other type of static storage device capable of storing static information and instructions, RAM or other type of dynamic storage device capable of storing information and instructions, or it may be an EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0080] The memory 5003 is used to store application code that executes the scheme of this application, and its execution is controlled by the processor 5001. The processor 5001 is used to execute the application code stored in the memory 5003 to implement the content shown in any of the foregoing method embodiments.

[0081] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers.

[0082] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a wind prediction method for power grid transmission channels as shown in the first aspect of this application.

[0083] Another embodiment of this application provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0084] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0085] In summary, this invention relates to the field of power system security protection technology, and discloses a method, system, equipment, and medium for predicting strong winds in power grid transmission channels. It constructs global meteorological features and local terrain features based on meteorological forecast data and geographic information data of the power grid transmission channel, respectively. These two types of features are then input into a composite neural network model based on convolutional neural networks and FiLM modulation for processing, outputting predicted wind field changes in the power grid transmission channel at a first-level spatial resolution. Based on real-time wind speed data of the power grid transmission channel, the initial wind field at a second-level spatial resolution is determined using a terrain-corrected dual-source distance-weighted interpolation method. The predicted changes at the predicted time are vector-synthesized with the initial wind values ​​to form a strong wind prediction result. This achieves refined strong wind forecasting with kilometer-level spatial resolution and short-term forecasts (1-24 hours) for power transmission channels in complex terrain areas.

[0086] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0087] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for predicting strong winds in power grid transmission channels, characterized in that, include: Real-time wind speed data, meteorological forecast data, and geographic information data of the power grid transmission channel are acquired, and global meteorological features and local terrain features are constructed based on the meteorological forecast data and the geographic information data, respectively. A composite neural network model is constructed based on convolutional neural network and FiLM modulation. The global meteorological features and the local terrain features are input into the composite neural network model for processing, and the predicted value of wind field change of the power grid transmission channel at the first-level spatial resolution is output. The real-time wind speed data is processed by a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-level spatial resolution. Based on the predicted wind field change and the wind field at the initial moment, the predicted wind conditions for the power grid transmission channel are determined.

2. The method for predicting strong winds in power grid transmission channels according to claim 1, characterized in that, The construction of global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data respectively includes: The center point coordinates and target area are determined based on the power grid transmission channel, and the upper-air meteorological area is set based on the center point coordinates and the target area; Extract geopotential height and temperature data for a preset time period from the meteorological forecast data; The geopotential height and temperature data are processed separately to obtain geopotential height difference and temperature difference, which are used as the global meteorological features.

3. The method for predicting strong winds in power grid transmission channels according to claim 2, characterized in that, The geographic information data includes elevation data and land use data; wherein... The construction of global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data respectively also includes: The elevation data is resampled to the target area using bilinear interpolation or cubic convolution to obtain ground elevation data. Based on the ground elevation data, the slopes in the east-west and north-south directions are calculated respectively to obtain the slope angle of the target area; The land use data is spatially resampled using the principle of maximizing area proportion to obtain the cover type of the target area, and the ground roughness is determined based on the cover type. The local terrain features are obtained by combining the ground elevation data, the slope angle, and the ground roughness.

4. The method for predicting strong winds in power grid transmission channels according to claim 2, characterized in that, The composite neural network model includes a CNN-based global context extractor, a CNN-based terrain feature extractor, a FiLM modulation layer, and an output decoder; wherein, The process of inputting the global meteorological features and the local terrain features into the composite neural network model for processing, and outputting the predicted wind field change value of the power grid transmission channel at a first-level spatial resolution, includes: The global meteorological features are extracted using the global context extractor to obtain a global context vector; The terrain feature extractor is used to extract features from the local terrain features to obtain a terrain feature map; Based on the FiLM modulation layer, the terrain feature map is subjected to dual conditional modulation using the global context vector to obtain modulation fusion features; The output decoder is used to progressively refine the modulation and fusion features, and outputs the predicted wind field change value of the power grid transmission channel at the first-level spatial resolution.

5. The method for predicting strong winds in power grid transmission channels according to claim 4, characterized in that, The step of using the global context vector to perform dual conditional modulation on the terrain feature map based on the FiLM modulation layer to obtain modulation fusion features includes: Based on the target region, a coarse adjustment parameter set is determined, and the global context vector is input to the first set of independent fully connected layers. After Tanh activation, the first scaling factor and the first offset factor under the coarse adjustment parameter set are generated. Upsample the first scaling factor and the first offset factor respectively to obtain a sampling scaling factor and a sampling offset factor with the same spatial resolution as the terrain feature map; The terrain feature map is scaled and offset channel by channel using the sampling scaling factor and the sampling offset factor to obtain a coarse-grained terrain feature map with global coarse adjustment. The global context vector is input into a second set of independent fully connected layers. After Tanh activation, a second scaling factor and a second offset factor with the same spatial resolution as the terrain feature map are generated. The coarse-grained terrain feature map is subjected to secondary channel-by-channel scaling and offset processing using the second scaling factor and the second sampling offset factor to obtain locally fine-tuned modulation and fusion features.

6. The method for predicting strong winds in power grid transmission channels according to claim 3, characterized in that, The real-time wind speed data is obtained through dual-source acquisition via a tower anemometer and a meteorological station; the real-time wind speed data includes first-source wind speed data and second-source wind speed data; wherein... The process of processing the real-time wind speed data using a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution includes: The target region is divided based on the second-level spatial resolution to obtain several target interpolation grids, and the target points in each target interpolation grid are determined. Calculate the horizontal spatial distance from each of the target points to any of the tower wind measuring devices or the meteorological station, and determine the distance weight of each of the horizontal spatial distances using the Gaussian attenuation rule; Based on the potential height, the ground elevation data and the slope angle, the terrain feature differences from each of the target points to any of the tower wind measuring devices or the meteorological station are calculated to obtain the comprehensive terrain similarity weight of each of the target points; Based on the distance weights and the comprehensive terrain similarity weights, the interpolation weights of each target point are determined to perform weighted interpolation processing on the first source wind speed data and the second source wind speed data to obtain the preliminary wind field values ​​of each target point at the second-level spatial resolution. The initial wind field values ​​are combined after being processed by terrain modulation and smoothing to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution.

7. A method for predicting strong winds in power grid transmission channels according to claim 6, characterized in that, The initial wind field values ​​are combined after terrain modulation and smoothing to obtain the initial wind field of the power grid transmission channel at a second-order spatial resolution, including: Based on the slope angle, the slope data of each target point is determined. Combined with the terrain position index that reflects the relative elevation of the target point, the east-west component of each preliminary wind field value is corrected to obtain the east-west corrected wind field of each target point. Based on the slope angle, the slope aspect data of each target point is determined. Combined with the actual wind direction angle in the real-time wind speed data, the north-south component in each preliminary wind field value is corrected to obtain the north-south corrected wind field of each target point. By combining the east-west and north-south corrected wind fields, the initial wind field of the power grid transmission channel at a second-order spatial resolution is obtained.

8. A high-wind forecasting system for power grid transmission channels, characterized in that, include: The data processing module is used to acquire real-time wind speed data, meteorological forecast data, and geographic information data of the power grid transmission channel, and to construct global meteorological features and local terrain features based on the meteorological forecast data and the geographic information data, respectively. The wind field prediction module is used to construct a composite neural network model based on convolutional neural network and FiLM modulation, and input the global meteorological features and the local terrain features into the composite neural network model for processing, and output the predicted value of wind field change of the power grid transmission channel at the first-level spatial resolution; The wind field construction module is used to process the real-time wind speed data using a terrain-corrected dual-source distance-weighted interpolation method to obtain the initial wind field of the power grid transmission channel at a second-level spatial resolution. The strong wind prediction module is used to determine the strong wind prediction result of the power grid transmission channel based on the wind field change prediction value and the wind field at the initial time.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the high wind prediction method for power grid transmission channels as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the wind prediction method for power grid transmission channels as described in any one of claims 1 to 7.

Images