Meteorological monitoring method for power grid equipment and meteorological forecast data mapping method

Abstract

The application discloses a kind of meteorological monitoring methods and meteorological forecast data mapping methods for power grid equipment.Mapping method includes: obtaining multi-source meteorological monitoring data and space-time alignment;Using multi-grid four-dimensional variation method, data and numerical model background field are fused from coarse to fine layer by layer, to generate grid analysis field;High-resolution grid field is output using deep residual network;With power grid equipment coordinates as center, determine the grid point covered by square window, calculate mapping weight and normalize weighted sum, obtain device location accurate meteorological monitoring value.Mapping method includes: using accurate monitoring value as initial constraint, extract the space-time characteristics of numerical prediction through space-time coding network;Combined with power grid equipment graph structure data, use attention mapping network to query device node, dynamically map adjacent prediction features as key and value, output device refined weather forecast value after gate fusion static attribute.The application is used to improve the accuracy of meteorological monitoring and forecasting of power grid equipment location.

Description

Technical Field

[0001] This invention belongs to the field of meteorological information technology, and in particular relates to a meteorological monitoring method for power grid equipment and a meteorological forecast data mapping method. Background Technology

[0002] Meteorological conditions are a key factor affecting the safe and stable operation of power grid equipment and power system dispatch decisions. Accurate equipment-level meteorological monitoring data is crucial for load forecasting, new energy power generation forecasting, disaster early warning (such as lightning strikes, wildfires, and icing), and equipment condition assessment. Currently, the main method for obtaining meteorological information on the location of power grid equipment relies on gridded analysis products published by professional meteorological agencies. However, the spatial resolution of such products is usually coarse (e.g., several kilometers to tens of kilometers), making it difficult to accurately reflect the micro-meteorological conditions at the location of specific power facilities such as substations and transmission line corridors. Directly assigning low-resolution grid data to equipment coordinates will introduce significant errors and cannot meet the needs of refined power grid operation. Although some spatial interpolation methods (such as nearest neighbor method and bilinear interpolation) exist to extrapolate grid data to specific points, these methods are based solely on mathematical relationships and do not fully consider the actual spatial distribution physical laws of meteorological elements under complex terrain and underlying surface conditions. Especially when power grid equipment is located in areas with drastic changes in meteorological gradients, the reliability of the interpolation results is limited. Therefore, how to generate high-precision data that accurately reflects the micro-meteorological characteristics of the location of power grid equipment based on multi-source meteorological observation data, i.e. how to improve the monitoring accuracy of meteorological elements at the location of power grid equipment, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a meteorological monitoring method and a meteorological forecast data mapping method for power grid equipment, which addresses the shortcomings of the prior art. By using a multi-grid four-dimensional variational method and a deep residual network to generate a high-resolution meteorological grid field, and combining it with adaptive mapping for the geographical coordinates of power grid equipment, the accuracy of meteorological monitoring of the location of power grid equipment is significantly improved.

[0004] The first aspect of this invention discloses a meteorological monitoring method for power grid equipment, comprising: Obtain a meteorological dataset within the target area. The meteorological dataset contains meteorological monitoring data from multiple sources. The meteorological monitoring data is spatiotemporally aligned using timestamps and coordinate information to obtain spatiotemporally unified multi-source observation data. The multi-grid four-dimensional variational method is used to refine the background field of the multi-source observation data and the numerical model from coarse to fine, and then gradually fused them to obtain a gridded analysis field with a predetermined spatial resolution and a predetermined temporal resolution. The multi-grid four-dimensional variational method is a cost function that includes a background term, an observation term, and a temporal increment smoothing term. A deep residual network is used to extract the low-resolution gridded analysis field and output the corresponding high-resolution grid field. Obtain the geographic coordinate information of the power grid equipment to be mapped within the target area, wherein the power grid equipment includes at least one of transformer substations, power distribution areas, and transmission lines; For each power grid device to be mapped, a set of grid points centered on the geographic coordinates of the power grid device is determined from the high-resolution grid field, and covered by a square window with a preset side length, which is used as the set of grid points in the spatial neighborhood of the power grid device. For each grid point within the neighborhood, the mapping weight of the grid point relative to the power grid equipment is calculated; based on the mapping weight of each grid point within the neighborhood, the meteorological element values ​​of each grid point are normalized and weighted to obtain the accurate meteorological element monitoring values ​​of the location of the power grid equipment.

[0005] The above method, wherein the spatiotemporal alignment is based on the coefficient of variation. An implementation of a cubic spline interpolation algorithm for adaptively adjusting window width, including: The adaptive window cubic spline interpolation algorithm is used to process discrete meteorological monitoring data, wherein the interpolation window width is... From the coefficient of variation of the data ; The adaptive window sliding interpolation algorithm uses cubic spline interpolation, and its interpolation process is expressed as follows:

[0006] In the formula, These are the interpolated meteorological element values ​​that change continuously over time. For the original discrete data point set, For timestamps, These are the meteorological element values ​​for the corresponding time. Based on the coefficient of variation Adaptive adjustment of the interpolation window width. Defined as , The standard deviation of the data points within the window. It is the mean, dimensionless.

[0007] The meteorological monitoring data from the multiple sources mentioned above include satellite remote sensing data, radar monitoring data, and ground observation data, wherein the ground observation data includes meteorological department station observation data and power grid micro-meteorological station observation data.

[0008] The cost function described above is:

[0009] Let be the state vector of the analysis field to be solved at the beginning of the time window, and be the initial condition of the analysis field at all subsequent time points; The background field is in its initial state. The background error covariance matrix describes the spatial correlation between grid points and variables in the background field. The number of data source categories observed; This represents the number of discrete time steps within the time window. For the first Observational data at time The observation vector; For a moment The observation operator maps the analysis field from the model space to the observation point; For a moment The observation error covariance matrix reflects the error characteristics of each observation instrument; For a moment The analytical field state is determined by It is obtained through time integration in a numerical model, i.e. , For model forecasting operators; This is a tangent linear mode, used to linearize the state evolution between adjacent time steps; Let be the mode error covariance matrix, which describes the uncertainty accumulated by the mode during time integration.

[0010] The above method, These correspond to satellite remote sensing data, radar monitoring data, ground meteorological station observation data, and power grid micro-meteorological station observation data, respectively. The time window is 3 hours, and the time resolution is 1 hour.

[0011] The above method employs a deep residual generator structure in the deep residual network. It extracts multi-scale features through multiple densely connected dilated convolutional blocks, introduces upsampling skip connections to retain large-scale background information, and uses a composite loss function during training to simultaneously optimize pixel-level mean square error, Laplacian smoothing term, and structural similarity index.

[0012] The downscaling mapping relationship of the deep residual network described above is as follows:

[0013] in, The output is a high-resolution meteorological field with dimensions of , The number of rows and columns of the high-resolution grid. Number of meteorological element channels; The input is a low-resolution meteorological field with dimension . upsampling factor ; For parameters The deep residual generative network, composed of densely connected dilated convolutional blocks, is used to learn high-frequency detail information; For bicubic interpolation upsampling operators, the output is the same as... Baseline fields of the same dimension.

[0014] In the above method, the mapping weights are calculated according to the following formula:

[0015] in, For power grid equipment points The combined meteorological element values; For equipment points around The set of all high-resolution grid point indices within the window; For grid points Weather values ​​at the location; These are the weight coefficients before normalization; the final weights are calculated by dividing by... Complete normalization; For equipment points With grid points The Euclidean distance between them; This is the distance attenuation scale parameter; This refers to the gradient alignment coefficient; Let be the cosine of the angle between the meteorological gradient direction and the direction of the line connecting the grid point to the equipment point, where , For grid points The meteorological gradient vector at that location, From grid points Pointing to device The vector.

[0016] The above method also includes an urban canyon effect correction factor in the calculation of the mapping weights. :

[0017] in, This represents the average height of buildings in the area. The average street width in the area. The depth-to-width ratio of the street canyon reflects the canyon effect; the larger the ratio, the more pronounced the canyon effect. The empirical correction coefficients for different types of built-up areas are determined by referring to a table based on the building height distribution density within the area. For wind speed forecasting using meteorological grids, For reference wind speed; The floor area of ​​the high-rise buildings in this area. The ratio represents the total area of ​​the region. Spatial distribution reflecting regional building density.

[0018] In the above method, the predetermined spatial resolution is 100 meters, the predetermined time resolution is 1 minute, and the preset side length is 5, corresponding to a square window with a range of 500 meters.

[0019] The meteorological element values ​​mentioned above include at least one of the following: 2m air temperature, 2m specific humidity, 10m wind speed, surface air pressure, and hourly precipitation.

[0020] A second aspect of this invention discloses a method for mapping meteorological forecast data for power grid equipment, comprising: Using the method described in claim 1, accurate meteorological element monitoring values ​​of the locations of each power grid device within the target area are obtained; Meteorological numerical forecast data of the target area within a preset forecast period is obtained. The precise meteorological element monitoring values ​​are used as initial state constraints. A spatiotemporal coding network is constructed to extract spatiotemporal features from the meteorological numerical forecast data to obtain forecast spatiotemporal features. Obtain graph structure data of power grid equipment within the target area, wherein the graph structure data includes the geographic coordinate information and static attribute information of each equipment node; An attention mapping network is used to dynamically map the device node as the query and the forecast spatiotemporal features within its spatial neighborhood as the key and value. The attention aggregation result is then fused with the static attribute information of the device node through a gating fusion mechanism to output the refined weather forecast value of the device node.

[0021] The above mapping method uses a spatiotemporal coding network, which is a three-dimensional convolutional spatiotemporal encoder composed of multiple stacked three-dimensional convolutional layers. Its abstract representation is as follows:

[0022] in, The input is a numerical weather forecast sequence with dimension [missing information]. ; For time steps, The number of rows and columns of the spatial grid. Number of meteorological element channels; This is a 3D convolutional encoder function, consisting of multiple convolutional layers, batch normalization layers, and ReLU activation functions stacked together; The output is a low-dimensional spatiotemporal feature tensor with dimension . ,in This represents the number of encoder output channels.

[0023] The above mapping method uses a multi-head cross-attention network for attention mapping. The gating fusion mechanism uses learnable gating coefficients to weight and combine dynamic features and static attributes. The core operation is represented as follows:

[0024] in, For device nodes Refined fusion of feature vectors; For nodes The initial forecast feature vector is obtained by bilinear sampling from Based on node coordinates Extracted; For nodes A set of spatially proximate forecast grid point indices; For neighboring grid points The predicted feature vector; For multi-head cross-attention function, For query, Given keys and values, output the aggregated dynamic features; For device nodes Static attribute encoding vector; It is a gated fusion function that uses learnable gating coefficients to weight and combine dynamic features and static attributes.

[0025] The above mapping method also incorporates a typhoon composite risk factor in the calculation of attention weights in the multi-head cross-attention network. :

[0026] When there is a typhoon impact, this factor should be included; among which... The weighting coefficients are dynamically adjusted based on the spatial relationship between the typhoon path and the urban coastal area. To forecast wind speed, The design wind speed for this transmission line; To forecast hourly rainfall intensity, Heavy rainfall warning thresholds set for urban power grids; To predict tide levels, This is the warning tide level for the foundations of coastal power poles.

[0027] The above mapping method includes a constraint term on the upper limit of urban heat island intensity in the training loss function of the attention mapping network:

[0028] in, This represents the heat island constraint loss value. This is a set of all urban equipment-suburban equipment pairings. Urban equipment is defined as a power grid node located in the central urban area of ​​the city, and suburban equipment is defined as a node located in the far suburbs of the city that belongs to the same meteorological forecast grid coverage area as urban equipment. and The urban equipment output by the model and suburban equipment Temperature forecast value; The upper limit threshold for the intensity of a typical urban heat island is given, and the loss term constrains the model output to satisfy... .

[0029] The above mapping method is characterized in that the training loss function further includes a narrowing effect wind speed correction loss term. :

[0030] in A collection of equipment nodes located in densely populated urban high-rise building areas; The wind speed value at this node represents the original weather forecast grid. These are the wind speed forecast values ​​after model correction; This is the correction factor for the aforementioned canyon effect.

[0031] Compared with existing technologies, this invention has the following advantages: By employing a multi-grid four-dimensional variational method to perform spatiotemporal alignment and physical constraint fusion of multi-source meteorological observation data, this invention generates a gridded analysis field with better physical consistency, laying the foundation for high-precision data. Subsequently, a deep residual network is innovatively introduced to intelligently reconstruct this analysis field using super-resolution, generating a grid field with higher spatial resolution capable of characterizing complex terrain and underlying surface influences, thus significantly improving the ability to describe micro-meteorological conditions at the data source. Finally, for the geographical coordinates of power grid equipment (such as distribution areas and substations), an adaptive weight mapping algorithm based on spatial proximity is designed to accurately and reasonably aggregate meteorological information from the high-resolution grid field to the specific locations of the equipment. This organic combination of technologies effectively overcomes the errors caused by direct assignment or simple mathematical interpolation of traditional low-resolution grid data, ultimately achieving a significant improvement in the monitoring accuracy of meteorological elements at the location of power grid equipment, providing more reliable data support for the refined operation and safety control of the power grid.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0034] Figure 2 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0035] Example 1 like Figure 1 As shown, a meteorological monitoring method for power grid equipment includes: Obtain a meteorological dataset within the target area. The meteorological dataset contains meteorological monitoring data from multiple sources. The meteorological monitoring data is spatiotemporally aligned using timestamps and coordinate information to obtain spatiotemporally unified multi-source observation data. The multi-grid four-dimensional variational method is used to refine the background field of the multi-source observation data and the numerical model from coarse to fine, and then gradually fused them to obtain a gridded analysis field with a predetermined spatial resolution and a predetermined temporal resolution. The multi-grid four-dimensional variational method is a cost function that includes a background term, an observation term, and a temporal increment smoothing term. A deep residual network is used to extract the low-resolution gridded analysis field and output the corresponding high-resolution grid field. Obtain the geographic coordinate information of the power grid equipment to be mapped within the target area, wherein the power grid equipment includes at least one of transformer substations, power distribution areas, and transmission lines; For each power grid device to be mapped, a set of grid points centered on the geographic coordinates of the power grid device is determined from the high-resolution grid field, and covered by a square window with a preset side length, which is used as the set of grid points in the spatial neighborhood of the power grid device. For each grid point within the neighborhood, the mapping weight of the grid point relative to the power grid equipment is calculated; based on the mapping weight of each grid point within the neighborhood, the meteorological element values ​​of each grid point are normalized and weighted to obtain the accurate meteorological element monitoring values ​​of the location of the power grid equipment.

[0036] In practice, taking the core business district of a coastal city as an example, the project covers a 110kV substation located near the city's main road, two overhead transmission lines along the main urban roads, and twenty-three distribution substations.

[0037] First, four types of meteorological monitoring data were acquired within the target area. The first type was geostationary meteorological satellite remote sensing data, with a temporal resolution of 15 minutes and a spatial resolution of approximately 4 kilometers, providing large-scale observational information such as cloud top brightness and temperature. The second type was monitoring data from two S-band Doppler weather radars within the urban area, with a temporal resolution of 6 minutes and a spatial resolution of approximately 1 kilometer, providing precipitation echo and wind field structure information. The third type was data from twelve national-level ground meteorological observation stations of the local meteorological department, with a temporal resolution of 1 minute. The fourth type was data from forty-six micro-meteorological stations deployed by the local power company on poles and transformer substations in and around the core business district, also with a temporal resolution of 1 minute. When aligning the above data spatiotemporally using timestamps and coordinate information, the timestamps of the ground observation data were used as a unified benchmark, and an adaptive window sliding cubic spline interpolation algorithm was employed to interpolate the satellite remote sensing and radar data. This algorithm first calculates the coefficient of variation of the data within the sliding window, i.e., the standard deviation divided by the mean. When the coefficient of variation is greater than 0.3, it indicates that the meteorological elements fluctuate drastically, such as the rapid change in cloud top brightness temperature when the outer cloud system of a typhoon passes through. The algorithm automatically narrows the interpolation window to three to five data points to preserve details. When the coefficient of variation is less than the threshold, it indicates that the element changes gradually, such as the slow decrease in surface air pressure before the arrival of a typhoon. The algorithm automatically widens the window to ten to fifteen data points to suppress observation noise.

[0038] After the aforementioned adaptive interpolation, satellite data was interpolated from a 15-minute interval and radar data from a 6-minute interval to a 5-minute time resolution. After time alignment, the geographic coordinates of all data were uniformly transformed to the same coordinate system, eliminating spatial discrepancies between different data sources, ultimately resulting in time-synchronized and spatially consistent multi-source observation data.

[0039] Next, a multi-grid four-dimensional variational method was employed to fuse multi-source observation data with the numerical model background field. The background field originated from the initial field of the regional numerical weather prediction model, with a spatial resolution of 0.125 degrees (approximately 14 kilometers). The multi-grid strategy progressively refined the spatial resolution from 14 kilometers to 1 kilometer. The first grid layer, at 14 kilometers, solved for large-scale incremental features; the second layer, at 7 kilometers, supplemented with mesoscale corrections; the third layer, at 3.5 kilometers; and the fourth layer, down to 1 kilometer, represented the final resolution. Each grid layer solved for the optimal analysis increment by minimizing the cost function, and the interpolated results from the previous layer served as the background constraint for the next layer. The time window was set to three hours, and the time step Nt was three, corresponding to each hour.

[0040] The cost function comprises three terms: the background term constrains the analysis field at the initial time to avoid deviating from the prior information of the background field through the background error covariance matrix B; the observation term reflects the measurement accuracy of each observation instrument through the observation error covariance matrix R, with satellite radiance observations having larger errors and thus lower weights, while ground standard station temperature and pressure observations having higher accuracy and thus higher weights; the observation operator H is responsible for mapping model grid variables to actual observation points; and the time increment smoothing term explicitly constrains the rate of change of the analysis field increment between adjacent time points through the tangent linear model. For example, when the outer wind area of ​​a typhoon moves towards the city at a speed of 30 kilometers per hour, this term requires that the wind field increment between each time point be consistent with the evolution trend predicted by the tangent linear model, preventing physically inconsistent jumps in the time series due to instantaneous fluctuations in observation data at individual time points.

[0041] By co-minimizing the triple cost function, the final output is a gridded analysis field with a spatial resolution of one kilometer and a temporal resolution of one hour, including elements such as near-surface air temperature at two meters, specific humidity at two meters, wind speed at ten meters, surface air pressure, and hourly precipitation.

[0042] Then, a deep residual network is used to downscale the gridded analysis field. A one-kilometer resolution cannot distinguish the local meteorological differences between the riverbanks running through the city, urban parks and green spaces, and dense high-rise building clusters in the core business district. Further downscaling is required to a spatial resolution of 100 meters and a temporal resolution of one minute.

[0043] The network employs a deep residual generator structure, consisting of multiple densely connected dilated convolutional blocks. Each convolutional block uses convolutional kernels with dilation rates of 1, 2, 3, and 5 for parallel processing. Small dilation rates capture local textures on the order of hundreds of meters, such as the subtle perturbations of the temperature field by a single block of green space, while large dilation rates capture background gradients on the order of kilometers, such as the overall temperature and humidity differences between riverside areas and the urban core. Dense connections between the convolutional blocks ensure efficient gradient propagation.

[0044] The network introduces upsampling skip connections to directly transmit large-scale background information from the encoder to the decoder, ensuring that large-scale meteorological features in the low-resolution field are not diluted during downscaling. The downscaling mapping relationship is: the high-resolution grid field equals the sum of the high-frequency detail information learned by the deep residual network and the baseline field obtained by bicubic interpolation upsampling. During training, the composite loss function simultaneously optimizes three terms: pixel-level mean square error, Laplacian smoothing term, and structural similarity index, respectively ensuring per-grid numerical accuracy, smoothness between adjacent grids, and spatial structure consistency. The final output is a high-resolution grid field with a spatial resolution of 100 meters and a temporal resolution of 1 minute.

[0045] Then, the precise geographic coordinates of each power grid device were extracted from the local power company's power grid geographic information system platform. Target substation A is located at approximately 121.5 degrees east longitude and 31.2 degrees north latitude, covering an area of ​​approximately 4,500 square meters. Line B is an overhead line extending north from substation A to the main urban road, with a length of approximately 3.2 kilometers and coordinates of 17 towers along the line accurate to the decimeter level. Line C runs south along another main urban road, with a length of approximately 4.1 kilometers and 21 towers along the line. In addition, 23 distribution transformer substations within the area were extracted and represented by their point coordinates. This established a spatial correlation between the spatial location of power grid devices and the high-resolution grid field.

[0046] For each power grid device to be mapped, a square window with a side length of 5 is defined in a high-resolution grid field, centered on the device's coordinates. This is achieved by taking five grid points along each of the latitude and longitude directions, forming a total of twenty-five 5x5 grid points. Since the grid resolution is 100 meters, this window covers an area approximately 500 meters squared around the device. Taking substation A as an example, the window extends northward to approximately 200 meters north of the substation, southward to approximately 200 meters south of the substation, and westward and eastward by approximately 200 meters as well. The same 5x5 window is determined for the towers on line B.

[0047] Finally, the mapping weights for each grid point within the window are calculated and summed using a weighted average. The weights consist of the product of a distance attenuation factor and a gradient alignment factor. The distance attenuation factor uses a Gaussian function with a scale parameter σ of 150 meters. When the grid point is zero distance from the equipment, the factor is close to its maximum value of one; at 150 meters, it attenuates to approximately 0.6; and at approximately 250 meters from the edge of the window, it attenuates to approximately 0.25, making closer grid points contribute more. The gradient alignment factor considers the cosine of the angle between the meteorological gradient direction at the grid point and the direction the grid point points towards the equipment. The gradient alignment coefficient η is 0.5. Taking substation A as an example, if at a certain moment there is a temperature gradient zone approximately 120 meters south of the substation, with a cooler region north of the gradient zone (shaded by tall buildings) and a region with ample sunshine to the south, the temperature gradient vector of the grid points on the gradient zone points towards the substation, and its gradient alignment factor is greater than one. The mapping weight is increased additionally on top of the distance attenuation factor. However, the gradient vector of the grid points east or west of the gradient zone is perpendicular to the direction pointing towards the substation, and the gradient alignment factor is close to one. The weight is determined solely by the distance factor. After normalizing the weights of all grid points within the window, the values ​​are weighted and summed with the meteorological element values ​​of each grid point to obtain the accurate meteorological element monitoring values ​​for that equipment point. The above process is performed one by one on all towers and twenty-three distribution areas of substation A, line B, and line C to complete the sub-grid precision meteorological mapping from a 100-meter grid field to discrete power grid equipment points.

[0048] In one embodiment, aligning the meteorological monitoring data spatiotemporally using timestamps includes: employing an adaptive window sliding cubic spline interpolation algorithm, based on the data's coefficient of variation. Dynamically adjust the width of the interpolation window; The adaptive window sliding interpolation algorithm uses cubic spline interpolation, and its interpolation process is expressed as follows:

[0049] In the formula, These are the interpolated meteorological element values ​​that change continuously over time. For the original discrete data point set, For timestamps, These are the meteorological element values ​​for the corresponding time. Based on the coefficient of variation Adaptive adjustment of the interpolation window width. Defined as , The standard deviation of the data points within the window. It is the mean, dimensionless.

[0050] In implementation, taking the cloud top brightness temperature data from a geostationary meteorological satellite in this embodiment as an example, the original time resolution of this data is fifteen minutes. Within the current time window, if the cloud top brightness temperature sequence is... ,in For each observation time, This represents the cloud top brightness temperature value at the corresponding time, in Kelvin. The algorithm first calculates the standard deviation of all data points within this window. and mean The coefficient of variation was obtained. .

[0051] When the outer convective clouds of a typhoon pass through, the cloud top brightness temperature can drop sharply from about 280 Kelvin to below about 200 Kelvin in a short period of time, with drastic fluctuations in the data. The value is relatively large, and Relatively less affected by extremely low values If the value exceeds the 0.3 threshold, the algorithm determines that the data is in a high-fluctuation state and automatically adjusts the interpolation window width. Narrowing the data to three to five points, equivalent to the 45 to 75 minute time range of the original data, a cubic spline function is used to interpolate the data points within this narrow window to obtain any target time within that time period. Continuous meteorological element values ​​at the location Cubic spline interpolation ensures the continuity of the first and second derivatives of the interpolation curve at the data points, avoiding the angle-breaking phenomenon at the connection between adjacent data segments in linear interpolation. Through narrow-window cubic spline interpolation, the detailed features of rapid changes in cloud top brightness temperature are preserved during periods of intense typhoon convection, avoiding the smoothing of extreme weather characteristics caused by wide-window smoothing.

[0052] When the target area is under clear skies before a typhoon arrives, the cloud top brightness temperature changes gradually, with very small fluctuations in data over several consecutive time periods. Approaching zero The value is much lower than the 0.3 threshold, so the algorithm automatically adjusts the interpolation window width. The window is widened to ten to fifteen data points, equivalent to a time range of 150 to 225 minutes. This wider window utilizes neighborhood information over a longer time span for cubic spline interpolation, effectively suppressing the impact of random noise from satellite sensors and instantaneous observation errors on the interpolation results, resulting in a smoother and more stable time-series curve. Similarly, for the original six-minute interval sequence of radar monitoring data, the algorithm adaptively adjusts the window based on the coefficient of variation of its reflectivity factor, using a narrow window in areas with strong precipitation echoes and a wide window in areas with weak echoes or clear skies.

[0053] By using an adaptive window sliding cubic spline interpolation algorithm, dynamic matching between the interpolation window width and the degree of data fluctuation is achieved. It retains the ability to distinguish details during periods of drastic weather changes and achieves smooth noise reduction during periods of stable weather, overcoming the shortcomings of fixed window interpolation in dealing with complex weather processes.

[0054] In one embodiment, the meteorological monitoring data from multiple sources includes satellite remote sensing data, radar monitoring data, and ground observation data, wherein the ground observation data includes meteorological department station observation data and power grid micro-meteorological station observation data.

[0055] In practice, satellite remote sensing data comes from the onboard radiometer of a geostationary meteorological satellite. Visible and infrared radiation information from the top of the atmosphere in the target area is acquired through scanning. After inversion, parameters such as cloud top brightness temperature, atmospheric precipitable water, and surface temperature can be obtained. In this embodiment, the satellite data has a temporal resolution of 15 minutes and a spatial resolution of approximately 4 kilometers. Its advantage lies in its observation range covering the entire target area and its surrounding hundreds of kilometers, enabling continuous tracking of the evolution of large-scale weather systems such as typhoons and severe convection. The main limitation of satellite data is its limited vertical resolution and weak ability to directly constrain near-surface meteorological elements.

[0056] The radar monitoring data comes from two S-band Doppler weather radars deployed in the target area. By emitting electromagnetic waves and receiving the backscattered echoes of precipitation particles, it acquires three-dimensional structural information such as precipitation intensity, radial wind field, and echo top height. The radar data has a time resolution of six minutes and a spatial resolution of approximately one kilometer. Its advantage lies in its ability to finely characterize the internal structure of precipitation systems, such as the distribution of strong precipitation cells within a typhoon's spiral rainband. The main limitations of radar data are that its detection range is restricted by terrain obstructions and the Earth's curvature, and it lacks the ability to directly observe meteorological elements in non-precipitation areas.

[0057] Meteorological observation data comes from twelve national-level surface meteorological observation stations deployed within the target area. These measurements are standardized according to World Meteorological Organization standards, providing near-surface air temperature (2 meters), specific humidity (2 meters), wind speed (10 meters), surface air pressure, and rain gauge precipitation with a time resolution of one minute. This type of data boasts high measurement accuracy and undergoes rigorous quality control, making it the most fundamental and binding observation source in the entire fusion system. However, its spatial distribution is sparse; the twelve stations are spaced approximately eight to ten kilometers apart within a 30-kilometer radius of the target area, making it difficult to characterize local meteorological differences within the city caused by building clusters, waterways, and green spaces.

[0058] The power grid micro-weather station observation data comes from 46 dedicated micro-weather stations independently deployed by the power company on poles and transformer substations in the core business district. Each station measures near-surface air temperature, humidity, wind speed, and air pressure with a time resolution of one minute. The micro-weather stations are located adjacent to power facilities, and their observations directly reflect the real local meteorological environment where the power grid equipment is located. For example, the wind speed observation data of a micro-weather station located on top of a pole among high-rise buildings in the core business district can reflect the actual wind load amplified at that location due to the building funneling effect. The four types of data are unified and fused using the adaptive interpolation and multi-grid four-dimensional variational method described in claim 1. Satellite and radar data compensate for the limited coverage of ground stations, meteorological department stations provide high-precision standardized benchmark constraints, and the power grid micro-weather stations anchor the fused results to the real local meteorological conditions where the power equipment is located, achieving full coverage of multi-level and multi-scale meteorological monitoring from space to ground.

[0059] In one embodiment, the cost function is:

[0060] Let be the state vector of the analysis field to be solved at the beginning of the time window, and be the initial condition of the analysis field at all subsequent time points; The background field is in its initial state. The background error covariance matrix describes the spatial correlation between grid points and variables in the background field. The number of data source categories observed; This represents the number of discrete time steps within the time window. For the first Observational data at time The observation vector; For a moment The observation operator maps the analysis field from the model space to the observation point; For a moment The observation error covariance matrix reflects the error characteristics of each observation instrument; For a moment The analytical field state is determined by It is obtained through time integration in a numerical model, i.e. , For model forecasting operators; This is a tangent linear mode, used to linearize the state evolution between adjacent time steps; Let be the mode error covariance matrix, which describes the uncertainty accumulated by the mode during time integration.

[0061] In practice, the first term of the cost function is the background term. Background error covariance matrix. The diagonal elements reflect the uncertainty level of each grid point and variable in the background field, while the off-diagonal elements reflect the error correlation between different grid points and different variables. For example, the background error of near-surface air temperature is usually positively correlated between two adjacent grid points, and the correlation is stronger the closer the distance. This spatial correlation structure is achieved through... The exponentially decaying correlation function in the matrix is ​​used to describe it. Deviation The further away, the greater the penalty for the background term, and the analysis field at the constraint start time cannot be completely separated from the prior information provided by the numerical model.

[0062] The second term of the cost function is the observation term, which cumulatively measures the observation bias of the four types of observation sources at all times within the time window. For satellite remote sensing data, the observation operator... The temperature and humidity profiles of the model layer are mapped to simulated brightness temperature values ​​at the top of the atmosphere using the radiative transfer equation, and then compared with the actual brightness temperature values ​​observed by satellite. Calculate the deviation, the deviation is... Weighted summation. Observation error covariance matrix. The larger the diagonal element, the lower the measurement accuracy of the observation instrument, and the smaller its weight in the assimilation. In this embodiment, the air pressure observation accuracy of the ground meteorological station is approximately 0.3 hPa. The diagonal elements have smaller values ​​and higher weight in the fusion process; the satellite brightness temperature observation accuracy is approximately one to two Kelvin. The diagonal elements have larger values ​​and relatively lower weights.

[0063] The third term of the cost function is a time-increment smoothing term. (Tangent linear mode) This describes the physical evolution relationship that the analytical fields should conform to between adjacent time points without additional observational constraints, given by a linearized form of the atmospheric kinematic equations. This penalty... and The difference between them requires that the temporal evolution of the analyzed field be consistent with the physical laws of atmospheric motion. Model error covariance matrix. The diagonal element represents the uncertainty of the tangent linear model in describing real atmospheric evolution. A larger diagonal element indicates lower model confidence at that moment and higher tolerance for deviations. During the rapid passage of typhoons, the approximation error of the tangent linear model increases. The corresponding values ​​are relatively high, allowing for some slack in the time increment smoothing term; under stable weather conditions, The lower value imposes stronger constraints on the continuity of the time series. By introducing a time increment smoothing term, the fusion results avoid non-physical jumps caused by observation noise or model errors in the time dimension, ensuring the physical consistency of the time series curves of meteorological elements.

[0064] In one embodiment, These correspond to satellite remote sensing data, radar monitoring data, ground meteorological station observation data, and power grid micro-meteorological station observation data, respectively. The time window is 3 hours, and the time resolution is 1 hour.

[0065] In implementation, this parameter setting is matched with the actual temporal characteristics of multi-source data and the operational needs of power grid meteorological monitoring. The four types of observation sources cover an altitude range from geostationary orbit (approximately 36,000 kilometers above the ground) to ground stations, complementing each other in terms of observation principles, spatial coverage, and error characteristics. The 3-hour time window setting comprehensively considers the evolution speed of the meteorological system and computational resources: if the window is too short, such as 1 hour, the amount of observational information is insufficient, and the assimilation results have a weak constraint on large-scale weather systems; if the window is too long, such as more than 6 hours, the approximation error of the tangent linear model will accumulate significantly, the constraint effect of the time increment smoothing term will decrease, and the computational load will increase superlinearly. The 3-hour window precisely covers the time span from the outer wind zone of the typhoon entering the target area to the core impact period, enabling the complete capture of the rapid evolution of the typhoon within a single assimilation window.

[0066] Time Steps This means discretizing consecutive time periods within a 3-hour time window at 1-hour intervals. There are four time points, corresponding to three time steps. At each time point, four types of observation data provide the observation constraints for that moment, while the time increment smoothing term constrains the physical consistency between three adjacent time pairs. This parameter configuration ensures sufficient observation information at each time step, while keeping the computational cost within an acceptable range, making it suitable for real-time or near-real-time power grid meteorological monitoring operations.

[0067] In one embodiment, the deep residual network adopts a deep residual generator structure, extracts multi-scale features through multiple densely connected dilated convolutional blocks, introduces upsampling skip connections to retain large-scale background information, and uses a composite loss function during training to simultaneously optimize pixel-level mean square error, Laplacian smoothing term, and structural similarity index.

[0068] In implementation, the core component of the depth residual generator is a densely connected dilated convolutional block. Each dilated convolutional block contains multiple parallel channels, each using a convolutional kernel with a different dilation rate to process the same input feature map. The dilation rate d controls the interval between adjacent sampling points in the convolutional kernel. When d equals 1, it degenerates into a standard convolution with a receptive field of 3x3; when d equals 3, the receptive field expands to 7x7; and when d equals 5, the receptive field further expands to 11x11. Multiple convolutional kernels with different dilation rates operate in parallel on the same layer, enabling a single-layer network to simultaneously perceive meteorological spatial features from local hectares to regional kilometers. For example, in a low-resolution gridded analysis field, the cooling effect of urban park green spaces is reflected at the hectare scale as a large temperature gradient between adjacent grids, while the large-scale temperature and humidity differences between riverside areas and inland areas are reflected across a span of ten to twenty grids. Different dilation rate channels of the same dilated convolutional block can capture features at these two scales respectively.

[0069] Dense connections refer to a method in which the output feature map of each dilated convolutional block in the generator is concatenated with the input of all subsequent convolutional blocks. This allows low-level features extracted from previous layers, such as edge and texture information, to be directly passed to deeper layers without the need for transformation and compression by intermediate layers. This connection method solves the gradient vanishing problem caused by increasing the number of layers in deep networks, enabling gradients to be directly propagated back to shallower layers during backpropagation through dense connections, thus accelerating network convergence. At the same time, dense connections promote feature reuse and reduce the number of parameters.

[0070] Upsampling skip connections refer to establishing direct cross-layer connections between the encoder and decoder paths in the generator. The encoder is responsible for progressively compressing spatial dimensions to extract high-level semantic features, while the decoder is responsible for progressively restoring spatial resolution to generate high-resolution output. Feature maps from each layer in the encoder are directly passed to the corresponding resolution layer in the decoder via skip connections, and then concatenated with the feature maps in the decoder. This ensures that large-scale structural information in the low-resolution background field preserved in the encoder—such as the overall location of the typhoon center and the banded distribution of the spiral rainbands—is not lost or distorted during the upsampling reconstruction in the decoder, avoiding the gradual blurring that occurs in traditional encoder-decoder networks as the number of upsampling layers increases.

[0071] The composite loss function during training consists of three weighted terms. The pixel-level mean square error loss calculates the mean of the squared differences between the downscaled output field and the true high-resolution observation field at each grid point, primarily constraining the numerical accuracy of the downscaling result to ensure that the meteorological element values ​​at each output grid point are as close as possible to the true values. The Laplace smoothing loss calculates the sum of the squares of the differences between the meteorological element values ​​at each grid point in the downscaled field and the mean values ​​of its adjacent grid points. This difference physically represents the local second derivative, constraining the downscaled field to have reasonable spatial smoothness and avoiding checkerboard artifacts such as abnormally prominent or concave values ​​at individual grid points. The structural similarity loss compares the similarity between the downscaled field and the true field in three dimensions: brightness, contrast, and structure. This comprehensively constrains the downscaling result to maintain the spatial structural characteristics of the true meteorological field, such as the curved shape of typhoon spiral rainbands and the linear gradient structure of the temperature field during cold front passage. These three loss functions are optimized synergistically during training, ensuring both the numerical accuracy at each grid point and the spatial physical rationality and structural fidelity of the field.

[0072] In one embodiment, the downscaling mapping relationship of the deep residual network is as follows:

[0073] in, The output is a high-resolution meteorological field with dimensions of , The number of rows and columns of the high-resolution grid. Number of meteorological element channels; The input is a low-resolution meteorological field with dimension . upsampling factor ; For parameters The deep residual generative network, composed of densely connected dilated convolutional blocks, is used to learn high-frequency detail information; For bicubic interpolation upsampling operators, the output is the same as... Baseline fields of the same dimension.

[0074] In practice, the core idea of ​​this mapping relationship lies in the residual learning paradigm. Bicubic interpolation upsampling operator. First, the low-resolution field Upsampling to high resolution via smooth interpolation, taking the downscaling from one kilometer to one hundred meters in this embodiment as an example, the upsampling factor... Bicubic interpolation uses a cubic polynomial weighting of sixteen neighboring low-resolution grid points around each target grid point to generate an upsampled field that already possesses the shape and size of a high-resolution grid and the basic spatial distribution of meteorological elements, i.e., the baseline field. However, the baseline field has an inherent smoothing defect: in areas where meteorological elements change drastically, such as the land-sea boundary, the boundary between urban and suburban areas, and the temperature edge of the shadow area of ​​high-rise buildings, bicubic interpolation cannot recover the sub-kilometer-scale high-frequency detail information lost in the one-kilometer resolution grid. The baseline field in these areas appears as a blurred transition zone.

[0075] Deep residual generation network The task is to learn and supplement these high-frequency detail corrections. Unlike directly learning the complete mapping from low to high resolution, residual learning focuses the network on learning the differences between the low-resolution and high-resolution fields—the residuals—while the large-scale background distribution is provided by the baseline field. Since the numerical magnitude of the high-frequency residuals is much smaller than the original field itself, the network's learning objective is significantly simplified, resulting in faster training convergence and better stability. For example, in a near-surface temperature field, the temperature at a grid point in a one-kilometer resolution field is 25 degrees Celsius, while in a 100-meter high-resolution field, the actual temperature distribution of the 10x10 subgrid in that region is between 24.5 and 25.8 degrees Celsius. The bicubic interpolation baseline field can only provide a uniform 25 degrees Celsius across all subgrids. Deep residual networks... The output residual field is then superimposed with corrections ranging from -0.5 to +0.8 degrees Celsius on each subgrid, resulting in the final... The area exhibits local temperature fluctuations consistent with the actual microclimate environment: the temperature is slightly lower on the shaded side of the building complex, slightly higher on the sunlit side and above the urban roads, and moderate in temperature along the riverbank due to the water body's moderating effect.

[0076] Through the downscaling mapping of this residual learning, a precise leap from a one-kilometer low-resolution field to a one-hundred-meter high-resolution field is achieved. This not only preserves the large-scale meteorological background information in the low-resolution field, but also uses a deep learning network to supplement the sub-kilometer-scale local meteorological details lost due to insufficient resolution.

[0077] In one embodiment, the mapping weights are calculated according to the following formula:

[0078] in, For power grid equipment points The combined meteorological element values; For equipment points around The set of all high-resolution grid point indices within the window; For grid points Weather values ​​at the location; These are the weight coefficients before normalization; the final weights are calculated by dividing by... Complete normalization; For equipment points With grid points The Euclidean distance between them; This is the distance attenuation scale parameter; The gradient alignment coefficient; Let be the cosine of the angle between the meteorological gradient direction and the direction of the line connecting the grid point to the equipment point, where , For grid points The meteorological gradient vector at that location, To from grid points Pointing to device The vector.

[0079] During implementation, power grid equipment points Integrated meteorological element values ​​at the location Meteorological values ​​of each grid point within its spatial vicinity We obtain the weighted summation as follows:

[0080] in For equipment points around The set of indices of all high-resolution grid points within the window. Unnormalized weighting coefficients. Proportional to:

[0081] in For equipment points With grid points The Euclidean distance between them, in meters; This is the distance attenuation scale parameter; The gradient alignment coefficient; Let be the cosine of the angle between the meteorological gradient direction and the direction of the line connecting the grid point to the equipment point. For grid points Meteorological gradient vector at location With from grid points Pointing to device vector The angle between them. The final weight is calculated by dividing by Normalization is complete.

[0082] In this embodiment, the device point Let's take substation A as an example. For each of the twenty-five grid points within its 5x5 window, calculate... , and The distance decay term adopts a Gaussian exponential decay form, and the scale parameter... Take 150 meters. At a certain grid point... It is located exactly at the location of substation A. When it is approximately zero, A value close to 1 indicates that the meteorological value at this grid point is highly correlated with the meteorological conditions at the equipment point, and should be assigned a weight close to the maximum; when the grid point is located in a corner of the window, Approximately 250 meters, calculated The weights decrease to about one-quarter. Gaussian attenuation has the advantage over linear or inverse distance attenuation in that the weights decrease more gradually in the vicinity of the device point, maintaining the balanced contribution of multiple grid points in close proximity; while in the far distance region, the weights decrease rapidly, effectively suppressing the undue influence of local noise on the fusion result of grid points at the window edge.

[0083] The gradient alignment term introduces the modulating effect of the meteorological gradient direction on the weights. Grid points Meteorological gradient vector at location This indicates the direction and intensity of the steepest change in meteorological element values ​​at that location. Taking near-surface air temperature as an example... The temperature gradient is approximated by the temperature difference between adjacent grid points in the longitude and latitude directions. If a grid point is located at the boundary between the shaded area and the sunlit area of ​​a tall building, its temperature gradient vector points in the direction of the fastest temperature increase, i.e., from the shaded side to the sunlit side. When this gradient vector happens to point to the location of the equipment... and Consistent direction Gradient alignment term close to 1 exist The maximum value of 1.5 is reached when the gradient vector is set to 0.5, indicating that the grid point is on the critical path of meteorological elements changing towards the equipment and should receive an additional weight boost. Conversely, if the gradient vector deviates from the equipment direction, A negative value indicates that the gradient alignment term is less than one, and the weights are suppressed. Gradient alignment coefficient. Controlling the gradient direction to modulate the intensity of the weights Taking 0.5 means that the additional contribution in the gradient direction can increase the weight by up to 50 percent or decrease it by the same amount.

[0084] Taking the specific operating conditions of substation A as an example. At a certain time during the outer edge of a typhoon, approximately 120 meters south of substation A lies the boundary between an urban park / green space and a cluster of high-rise commercial buildings. There is a temperature difference of about two to three degrees Celsius between the north and south sides of this boundary: the green space is cooler due to evaporation, while the hard paving between the commercial buildings and the heat dissipated by air conditioning units result in a higher temperature. A grid point on the boundary... Its temperature gradient vector Pointing due north means heading towards substation A. It also points due north. Approximately equal to one, the gradient alignment term is approximately 1.5, and this grid point has an additional unnormalized weight of approximately 1.05 on top of the distance decay weight of approximately 0.7. Another grid point on the east side of the boundary... Located within the green space, the temperature gradient vector runs almost east-west along the park boundary, and the angle between it and the vector pointing towards substation A is close to ninety degrees. The gradient alignment term is approximately zero, and the weight of this grid point is determined solely by the distance factor, remaining constant. After normalization, the weights of the final twenty-five grid points show that the grid points along the boundary zone with the temperature gradient pointing towards substation A have significantly higher weights than those in other directions, resulting in a more favorable fusion outcome. The data fully reflects the gradient characteristics of this temperature boundary zone, and the temperature monitoring value at the location of substation A more accurately reflects the combined effects of the cooling effect of the green space to its south and the warming effect of the commercial building complex.

[0085] By designing a mapping weight that simultaneously considers distance decay and gradient alignment, a refined quantitative expression of the spatial distribution characteristics of the meteorological field around the equipment is achieved, overcoming the shortcomings of traditional inverse distance weighted interpolation, which relies solely on distance and ignores the spatial variation trend of meteorological elements.

[0086] In one embodiment, the mapping weight calculation also includes an urban canyon effect correction factor. :

[0087] in, This represents the average height of buildings in the area. The average street width in the area. The depth-to-width ratio of the street canyon reflects the canyon effect; the larger the ratio, the more pronounced the canyon effect. The empirical correction coefficients for different types of built-up areas are determined by referring to a table based on the building height distribution density within the area. For wind speed forecasting using meteorological grids, For reference wind speed; The floor area of ​​the high-rise buildings in this area. The ratio represents the total area of ​​the region. Spatial distribution reflecting regional building density.

[0088] In implementation, the physical mechanism and calculation process of this factor will be illustrated using the core business district where substation A is located as an example. This area mainly consists of high-rise office buildings with an average building height of [missing information]. Approximately 80 meters wide, the main street is [width missing]. Approximately thirty meters, depth to width ratio Approximately 2.67. When the prevailing wind direction is parallel to the street direction, the airflow forms a significant tube acceleration effect between the tall buildings on both sides of the street: the incoming airflow is forced to converge into the relatively narrow street space due to the obstruction of buildings. According to Bernoulli's principle and the law of conservation of mass, the cross-sectional contraction leads to an increase in air velocity. Field studies show that in streets with a depth-to-width ratio greater than 2, the near-ground wind speed can be increased by 30% to 60% compared to the wind speed in open areas. This area is dominated by high-rise commercial buildings and is a high-rise dense area. An empirical correction factor was determined based on the building height distribution density table. Take 0.35. The total area occupied by high-rise buildings exceeding 50 meters in height within this region is approximately 0.12 square kilometers. The total area of ​​the region is approximately 0.3 square kilometers. The value is approximately 0.4, reflecting a high degree of high-rise building coverage in the area. The average summer wind speed in the region is approximately five meters per second.

[0089] When the outer bands of a typhoon are affected, the average forecast wind speed of the meteorological grid in that region is... Approximately twelve meters per second. Substituting into the formula, we can calculate:

[0090] The correction factor is approximately 1.9, meaning that under these specific conditions, the actual wind speed at the device's location may be about 1.9 times the grid-predicted wind speed. In the mapping weight calculation, this correction factor applies to the composite wind speed value at the device point: the original weighted summation wind speed value multiplied by... The wind speed value after correction for the canyon effect is obtained. If the uncorrected combined wind speed at substation A is 11.5 meters per second, the corrected wind speed is approximately 21.9 meters per second, which is significantly higher than the result without considering the canyon effect.

[0091] The design of this correction factor takes into account four dimensions of urban morphological parameters: the average building height through... The geometric constraints of street space are reflected through Reflected by building density This is reflected in the differences between built-up area types through Reflected by. Empirical correction coefficient. By looking up values ​​in a pre-established building morphology database, different values ​​correspond to different types of built-up areas: higher values ​​for densely built-up areas, intermediate values ​​for multi-story areas, and values ​​close to zero (i.e., almost no canyon effect) for sparsely built-up areas. By introducing an urban canyon effect correction factor, a refined expression of the urban micro-meteorological environment based on the meteorological mapping results of power equipment is achieved. This allows wind speed monitoring values ​​in densely built-up areas to reflect the local wind speed amplification caused by the funnel effect, avoiding a systematic underestimation of wind loads due to insufficient grid resolution to distinguish street-scale flow field structures.

[0092] In one embodiment, the predetermined spatial resolution is 100 meters, the predetermined time resolution is 1 minute, and the preset side length is 5, corresponding to a square window with a range of 500 meters.

[0093] In implementation, these parameter values ​​are a balanced choice made for the meteorological monitoring scenario of power grid equipment. The spatial resolution of 100 meters is based on the following considerations: the area occupied by a substation in the core business district is approximately 4,000 to 5,000 square meters, which is approximately 200 meters by 200 meters; the spacing between transmission line towers is usually between 100 and 300 meters; and the service radius of a distribution substation is approximately 200 to 300 meters. A high-resolution grid field with a resolution of 100 meters can place approximately four grid points within the substation area and one to three grid points between two adjacent towers, providing a sufficiently detailed meteorological background for each equipment mapping. If the resolution is further refined to 50 meters or smaller, the training difficulty and inference time of the downscaling network will increase significantly, while the marginal benefit for improving the accuracy of meteorological mapping of power grid equipment is limited. The temporal resolution of one minute enables the meteorological monitoring time series to capture short-term meteorological fluctuations at the minute level, such as the rapid changes in wind speed within a few minutes when a gust front passes, and the minute-level evolution of precipitation from strong convective cells, meeting the timeliness requirements for dynamic load assessment and short-term early warning of power grid equipment.

[0094] A preset side length L of 5 (5x5 window) corresponds to a square coverage area of ​​approximately 500 meters east-west and 500 meters north-south, centered on the device, under a spatial resolution of 100 meters. This window size balances local representativeness, statistical robustness, and computational efficiency. If the window is too small, such as 3x3 (300 meters), the number of neighboring grid points is only nine, and individual abnormal grid point values ​​may cause excessive disturbance to the weighted fusion results, resulting in insufficient statistical robustness. If the window is too large, such as 9x9 (900 meters), the neighborhood coverage is too large, which may introduce distant grid points with weak correlation to the device's micro-meteorological environment or even controlled by different local factors, thus diluting the contribution of truly valuable nearby grid points. The 500-meter range corresponding to the 5x5 window precisely covers the spatial scale of one or two complete blocks at the core business district scale. Within this range, urban micro-meteorological influencing factors such as building form, underlying surface properties, and street orientation are relatively uniform, and the neighboring meteorological elements have a reasonable spatial correlation with the meteorological conditions of the device.

[0095] In one embodiment, the meteorological element values ​​include at least one of 2m air temperature, 2m specific humidity, 10m wind speed, surface air pressure, and hourly precipitation.

[0096] In practice, the two-meter air temperature refers to the air temperature at a height of two meters above the ground, which is a fundamental parameter characterizing the near-surface thermal environment. For substations, ambient temperature directly affects the temperature rise of transformer oil and the rate of performance degradation of insulation materials; for overhead transmission lines, temperature changes cause thermal expansion and contraction of conductors, altering sag and thus affecting the safe distance to the ground. During a typhoon, the near-surface air temperature in the target area can plummet from approximately 33 degrees Celsius before the typhoon's arrival to around 25 degrees Celsius during the typhoon's impact.

[0097] Two-meter specific humidity refers to the mass of water vapor per unit mass of moist air at a height of two meters above the ground, reflecting the humidity level of the air near the ground. The combination of high specific humidity and high temperature conditions accelerates the electrochemical corrosion process of metal components. At the same time, high humidity reduces the air insulation strength and increases the risk of surface discharge of equipment.

[0098] Ten-meter wind speed refers to the horizontal wind speed at a height of ten meters above the ground, and it is the main wind load parameter acting on transmission towers and overhead conductors. Under typhoon conditions, ten-meter wind speeds can reach 25 to 35 meters per second or even higher, which has a decisive impact on the structural strength of towers and conductor galloping.

[0099] Surface air pressure refers to the atmospheric pressure near the ground. The low-pressure center of a typhoon can drop to about 960 to 980 hPa. The rapid drop in air pressure and spatial gradient are important indicators for judging the intensity and path of a typhoon, and are also reference parameters for monitoring the sealing performance of air-filled equipment such as GIS switchgear in substations.

[0100] Hourly precipitation refers to the cumulative rainfall on the ground over the past hour, which is directly related to the drainage capacity of substation sites, the risk of water accumulation in cable trenches, and the risk of erosion of transmission line tower foundations. During the heavy rainfall phase of a typhoon, hourly precipitation can reach 30 to 50 millimeters or more, posing a severe challenge to the flood control and drainage of power facilities.

[0101] The five types of elements are processed as independent channels in the downscaling and equipment mapping process, sharing the same spatiotemporal alignment, data fusion, deep downscaling and spatial proximity weighted mapping process, and finally obtaining accurate meteorological monitoring values ​​containing all five types of elements at each power grid equipment point.

[0102] Example 2 like Figure 2 As shown, a method for mapping meteorological forecast data for power grid equipment includes: Using the method described in Example 1, accurate meteorological element monitoring values ​​of the locations of each power grid device within the target area are obtained; Meteorological numerical forecast data of the target area within a preset forecast period is obtained. The precise meteorological element monitoring values ​​are used as initial state constraints. A spatiotemporal coding network is constructed to extract spatiotemporal features from the meteorological numerical forecast data to obtain forecast spatiotemporal features. Obtain graph structure data of power grid equipment within the target area, wherein the graph structure data includes the geographic coordinate information and static attribute information of each equipment node; An attention mapping network is used to dynamically map the device node as the query and the forecast spatiotemporal features within its spatial neighborhood as the key and value. The attention aggregation result is then fused with the static attribute information of the device node through a gating fusion mechanism to output the refined weather forecast value of the device node.

[0103] In practice, the technical solution is explained using a forecast lead time of six to seventy-two hours after the typhoon passes as an example. This embodiment provides a refined forecast of two-meter air temperature, ten-meter wind speed, and hourly precipitation for each tower of the target substation A, line B, and line C, as well as twenty-three distribution substations, every fifteen minutes within the forecast lead time.

[0104] First, the precise meteorological element monitoring values ​​of the locations of each power grid device within the target area are obtained using the method described in claim 1, serving as the current device-level meteorological data. This real-time data achieves device-level spatial resolution and minute-level temporal resolution. Compared to the traditional method of using meteorological station observations or model grid values ​​as the initial forecast field, it has the advantages of being completely consistent with the spatial location of the forecast object and being highly real-time in terms of time.

[0105] Then, meteorological numerical forecast data for the target area within the preset forecast lead time are obtained. In this embodiment, the forecast data comes from the hourly, three-kilometer resolution forecast field output by the regional high-resolution numerical forecast model, with a forecast lead time of seventy-two hours, and includes a three-dimensional forecast sequence of elements such as wind speed u-component, v-component, temperature, specific humidity, air pressure, and precipitation.

[0106] A spatiotemporal coding network is constructed to extract spatiotemporal features from the aforementioned meteorological numerical forecast data. This is done using the input meteorological numerical forecast sequence. The encoder input has the following dimensions: ,in The forecast time step is set to seventy-two. and The number of rows and columns in the spatial grid corresponds to a resolution of three kilometers. This represents the number of meteorological element channels. The spatiotemporal coding network... Convolutional encoding is performed, maintaining temporal continuity with a stride of 1 in the temporal dimension and progressively compressing resolution with a stride of 2 in the spatial dimension. After stacking multiple convolutional layers, a low-dimensional spatiotemporal feature tensor is output. This tensor retains the compressed spatial features at each time step, forming a condensed spatiotemporal feature representation of weather forecasts. Through a spatiotemporal coding network, the original high-dimensional numerical forecast data is compressed into a compact feature tensor containing deep spatiotemporal evolution patterns, enabling subsequent attention mapping networks to efficiently retrieve and aggregate data in a low-dimensional feature space containing sufficient forecast information.

[0107] Simultaneously, graph structure data of power grid equipment within the target area is acquired. The graph structure data uses power grid equipment as nodes, and each node contains two types of attributes: the first is geographic coordinate information, namely the precise latitude and longitude coordinates of each equipment node, which allows the equipment node to be located in the spatiotemporal feature tensor. The first category is the corresponding spatial location; the second category is static attribute information, including equipment type such as substation, tower or distribution area, rated capacity, voltage level, altitude, and surrounding building type. The topological connection relationship between equipment, such as the connection relationship between substation and line, and the series connection relationship of towers along the line, constitutes the edges of the graph, reflecting the physical connection and electrical association between equipment.

[0108] Then, an attention mapping network is used to dynamically map the spatiotemporal features. For each device node... First, its latitude and longitude coordinates are used to obtain the spatiotemporal feature tensor through bilinear sampling. An initial forecast feature vector corresponding to the node's location is extracted. This feature vector contains coarse-resolution forecast information for that location at various times within the forecast lead time. Then, using this initial forecast feature vector as the query and the spatiotemporal feature vectors of forecast grid points within a spatial neighborhood of the node (approximately a five-kilometer radius) as keys and values, attention calculation is performed. The attention mechanism assigns a greater aggregation weight to the value corresponding to the key with the highest similarity to the query. Essentially, it allows the device node to automatically find the forecast grid points most relevant to its location's meteorological conditions and richest in information within its spatial neighborhood, aggregating their forecast features. The dynamic features aggregated through attention are then fused with the static attribute information of the device node through a gating fusion mechanism. This gating fusion mechanism includes a learnable gating coefficient, which adaptively determines the proportion of dynamic meteorological features and static device attributes in the final fused features based on different device types and meteorological conditions. For example, for power poles located in open areas, wind speed forecasts mainly rely on dynamic meteorological characteristics, and gating assigns higher fusion weights to dynamic characteristics. Conversely, for power distribution stations surrounded by buildings, temperature and wind speed are significantly affected by the static attributes of the surrounding building forms, and gating assigns higher fusion weights to static attributes. The fused features are then output through a feedforward network as refined meteorological forecast values ​​for each time interval within the forecast period for that device node.

[0109] Taking substation A as an example, its spatial radius of five kilometers includes an open riverbank area approximately three kilometers to its southeast and a densely built-up area of ​​high-rise buildings approximately two kilometers to its northwest. Forecast grid points in different directions carry meteorological forecast signals of different natures: the wind speed forecast values ​​of the open riverbank area are less affected by topographic friction and tend to reflect wind speed levels in the free atmosphere; the temperature and wind speed of the forecast grid points in the densely built-up area are corrected for by the urban canopy effect. Through automatic learning, the attention mapping network adaptively assigns higher attention weights to the forecast grid points in the densely built-up area to the west, which has a more similar topography and building form to substation A's location, while only using a reference aggregation for the wind speed characteristics of the eastern riverbank area. The gating fusion mechanism further adjusts the fusion weights for temperature corrections and funnel effect wind speed corrections related to building form, based on substation A's static attributes of being a semi-indoor substation and its surrounding high-rise building form.

[0110] Through the above mapping process, the intelligent and refined forecast conversion from macroscopic three-kilometer resolution numerical forecast to microscopic power grid equipment level is realized. The forecast value of each equipment node not only absorbs the macroscopic grasp of weather system evolution from large-scale meteorological models, but also captures the personalized characteristics of its local micro-meteorological environment through the attention mechanism, and incorporates the correction effect of the equipment's own attributes on meteorological elements through gating fusion.

[0111] In one embodiment, the spatiotemporal coding network is a three-dimensional convolutional spatiotemporal encoder, composed of multiple stacked three-dimensional convolutional layers, which can be abstractly represented as follows:

[0112] in, The input is a numerical weather forecast sequence with dimension [missing information]. ; For time steps, The number of rows and columns of the spatial grid. Number of meteorological element channels; This is a 3D convolutional encoder function, consisting of multiple convolutional layers, batch normalization layers, and ReLU activation functions stacked together; The output is a low-dimensional spatiotemporal feature tensor with dimension . ,in This represents the number of encoder output channels.

[0113] In implementation, the key difference between 3D convolution and traditional 2D convolution lies in the fact that the convolution kernel slides simultaneously across three dimensions: time, spatial height, and spatial width. In this embodiment, the kernel size of each convolutional layer is... This means that it spans three consecutive time intervals in the time dimension and a 3x3 neighborhood grid in the spatial dimension. The convolution stride in the spatial dimension is set to 2 to halve the spatial size layer by layer, and the stride in the time dimension is set to 1 to maintain the temporal resolution.

[0114] Input in this embodiment Taking a forecast sequence of 72 time intervals, each covering approximately a 60x60 grid, as an example, the first 3D convolutional layer uses 32... The convolutional kernels, after convolution with a stride of 2 in the spatial dimension, reduce the spatial size from 60x60 to 30x30, while maintaining a constant 72 time intervals, increasing the number of output channels to 32. A batch normalization layer standardizes the features of each channel to accelerate training convergence, and a ReLU activation function introduces non-linearity to enhance the network's expressive power. The second layer uses 64 convolutional kernels, further reducing the spatial size to 15x15, while increasing the number of channels to 64. The third layer uses 128 convolutional kernels, reducing the spatial size to approximately 8x8. Number of channels The value is 128. In the three-dimensional convolution operation of each layer, the three consecutive time intervals spanned by the convolution kernel in the time dimension enable the encoder to learn the evolution of meteorological elements between adjacent time intervals, such as the continuous cooling trend of the temperature field in the three hours before and after the passage of the cold front, and the continuous enhancement of the wind speed field in the three hours during the approach of the typhoon. These temporal evolution patterns are encoded as high-dimensional representations in the spatiotemporal feature tensor.

[0115] By using a three-dimensional convolutional spatiotemporal encoder, joint feature extraction of meteorological numerical forecast sequences in both time and space dimensions was achieved. The original high-dimensional sparse forecast data was compressed into a compact feature representation rich in spatiotemporal evolution information, providing high-information-density and low-noise input features for subsequent attention mapping networks.

[0116] In one embodiment, the attention mapping network is a multi-head cross-attention network, and the gating fusion mechanism uses learnable gating coefficients to weight and combine dynamic features and static attributes. The core operation is represented as follows:

[0117] in, For device nodes Refined fusion feature vectors; For nodes The initial forecast feature vector is obtained by bilinear sampling from Based on node coordinates Extracted; For nodes A set of spatially adjacent forecast grid point indices; For neighboring grid points The predicted feature vector; For multi-head cross-attention function, For query, Given keys and values, output the aggregated dynamic features; For device nodes Static attribute encoding vector; It is a gated fusion function that uses learnable gating coefficients to weight and combine dynamic features and static attributes.

[0118] During implementation, firstly The features of each neighboring grid point are mapped to a query vector through a linear transformation. The query is mapped to key and value vectors through another set of linear transformations. The dot product of the query and each key is scaled and normalized using Softmax to obtain the attention weight distribution. The weighted sum of the value vectors yields the aggregated output of the single-head attention. The multi-head mechanism will... and The data is mapped to multiple attention heads for parallel computation through various linear transformations. Each head can focus on different spatial neighborhood patterns: for example, one head might focus on grid points with similar altitudes to the node location, as these grid points are more correlated with the device's temperature forecast; another head might focus on grid points upstream of the prevailing wind direction, as these grid points provide stronger early indication for the device's wind speed forecast. The outputs of the multiple heads are concatenated and linearly transformed to obtain the final attention aggregation dynamic feature.

[0119] Taking substation A as an example, there are approximately twenty-eight forecast grid points with a resolution of three kilometers within a five-kilometer radius of its spatial vicinity. When substation A queries these grid points, one focus of attention may assign higher weight to several grid points in the northwest direction, i.e., the upwind direction, because the prevailing wind direction during typhoon weather is due north-northwest, and the wind speed characteristics of these grid points are indicative of the wind speed evolution of substation A in the next few hours; another focus of attention may assign higher weight to several grid points at an altitude of about ten meters similar to that of substation A, because the temperature forecasts of these grid points are more thermodynamically correlated with the temperature forecast of substation A.

[0120] Gated fusion function The core is a learnable gating coefficient vector with the same dimension as the fused feature space. After activation by the sigmoid function, this gating coefficient takes values ​​between zero and one, representing the proportional weight of dynamic meteorological features in the final fused features. Final fused features The threshold is equal to the element-wise product of the gating coefficient and the dynamic features, plus one minus the element-wise product of the gating coefficient and the static attribute features. For wind speed forecasting, substation A is located in a densely populated area of ​​high-rise buildings, and its wind speed is significantly affected by the funneling effect of the surrounding building shapes. During training, the gating coefficient learns to give higher weight to the building shape-related features in the static attributes, resulting in a lower value for the gating coefficient in this dimension. This makes the influence of static attributes account for a larger proportion in the final wind speed forecast. For temperature forecasting, the temperature of substation A is mainly controlled by large-scale weather systems and solar radiation, with relatively less influence from static attributes. Therefore, the gating coefficient is set higher, and dynamic meteorological features dominate.

[0121] By combining multi-head cross-attention with gating fusion mechanism, intelligent information aggregation and personalized fusion from macro forecast grid to micro device node are realized. The forecast value of each device node not only makes full use of the meteorological forecast information of its spatial neighborhood, but also incorporates the correction effects of the device's own attributes and local environmental characteristics.

[0122] In one embodiment, a typhoon composite risk factor is also introduced into the attention weight calculation of the multi-head cross-attention network. :

[0123] When there is a typhoon impact, this factor should be included; among which... The weighting coefficients are dynamically adjusted based on the spatial relationship between the typhoon path and the urban coastal area. To forecast wind speed, The design wind speed for this transmission line; To forecast hourly rainfall intensity, Heavy rainfall warning thresholds set for urban power grids; To predict tide levels, This is the warning tide level for the foundations of coastal power towers.

[0124] This factor is triggered when the typhoon's path is less than 200 kilometers from the target city's coastline. (Three weighting coefficients) , , Dynamic adjustments are made based on the spatial relationship between the typhoon's path and the target area: When the typhoon center is located approximately 150 kilometers east of the target city at sea, and the typhoon's direction of movement is due north-west, the target city is located to the right of the typhoon's direction of movement, i.e., in the danger semi-circle area, where wind speeds are at their maximum and the city is directly opposite the coastline. Take 0.6 Take 0.25 Taking 0.15, the risk weight for wind-induced disasters is the highest; if the typhoon's path is southerly, brushing past the target city about 200 kilometers south and moving westward, the target city will mainly be affected by peripheral rainfall. The weight is increased to 0.4. Dropped to 0.4 It dropped to 0.2.

[0125] Take a power pole located in a coastal area on line C as an example. The design wind speed of this power pole is... The wind speed is 35 meters per second, and the affected area is a coastal area with strong winds. When the typhoon center is located approximately 150 kilometers east of the target city, the forecast wind speed is... The forecast is 28 meters per second, representing the hourly rainfall intensity. The power company set a heavy rainfall warning threshold of 40 millimeters per hour. The forecast tide level is 30 millimeters per hour. The warning tide level for the tower foundation is 4.2 meters. It is 4.5 meters. Substitute into the calculation:

[0126]

[0127]

[0128] The composite risk factor of approximately 0.86 indicates that the tower is simultaneously subjected to high wind loads and a combined threat of near-warning-level rainfall intensity and tide levels under current typhoon conditions. This factor is embedded in the attention weight calculation, enabling the network to assign additional attention sensitivity to equipment nodes in high-risk states and grid points within forecast danger zones when aggregating spatial neighborhood forecast features: when the composite risk factor value of an equipment node is high, the network will more finely search for forecast features reflecting extreme weather conditions in its spatially neighboring grid points, improving the weather forecast resolution and response strength for equipment in high-risk areas. If at the same time another tower is located inland, with a forecast wind speed of only 18 meters per second, rainfall intensity of 15 millimeters per hour, and unaffected by tide levels, its... Well below 0.86, the prediction aggregation pattern of the attention network will be relatively stable.

[0129] By introducing the combined risk factors of typhoon, wind, rain and tide, the forecasting model can simultaneously assess the multi-factor coupled threat of wind load, rain erosion and tide rise to power grid equipment under combined severe weather conditions such as typhoons. It can accurately identify highly vulnerable sections of transmission lines that are attacked by multiple disaster factors at the same time, and provide quantitative and refined forecasting support for flood control and typhoon response scheduling of coastal power grids.

[0130] In one embodiment, the training loss function of the attention mapping network includes an upper bound constraint on the intensity of the urban heat island:

[0131] in, This represents the heat island constraint loss value. This is a set of all urban equipment-suburban equipment pairings. Urban equipment is defined as a power grid node located in the central urban area of ​​the city, and suburban equipment is defined as a node located in the far suburbs of the city that belongs to the same meteorological forecast grid coverage area as urban equipment. and The urban equipment output by the model and suburban equipment Temperature forecast value; The upper limit threshold for the intensity of a typical urban heat island is given, and the loss term constrains the model output to satisfy... This means that the temperature difference between equipment in urban areas and equipment in suburban areas should not exceed the upper limit of heat island intensity.

[0132] During implementation, the urban equipment is located at Substation A and eight distribution substations within the core business district, while the suburban equipment is located at four distribution substations in the far suburbs, approximately 30 to 40 kilometers from the core business district. All of the aforementioned urban and suburban equipment are within the coverage area of ​​the same meteorological forecast grid, meaning they are controlled by the same solar radiation background, large-scale weather patterns, and seasonal climate conditions. The temperature difference between the two mainly stems from the urban heat island effect. The temperature was set at 4.0 degrees Celsius, determined based on the average maximum temperature difference between urban and suburban areas, calculated from daily meteorological observation data of the region over the past ten years.

[0133] During model training, when a batch of training samples includes a single forecast case under clear, calm nighttime conditions, the urban heat island effect is most pronounced under these conditions: the heat absorbed and stored in buildings and roads during the day in the core business district is slowly released at night, coupled with the continuous input of heat from air conditioning, causing the nighttime temperature in the urban area to remain consistently higher than that in the suburbs. If the model, due to overfitting to a few abnormally high-temperature records in the training data, outputs a result showing an urban equipment temperature of 35 degrees Celsius and a suburban equipment temperature of 29.5 degrees Celsius on such samples, the temperature difference is 5.5 degrees Celsius, exceeding [a certain threshold]. If the upper limit is 4.0 degrees Celsius, then the pairing will contribute a loss value of 5.5 minus 4.0 equaling 1.5. In the process, backpropagation guides the model parameters to adjust in a direction that reduces the excessive temperature difference. In a forecast example of cloudy weather during the day, the temperature on the urban equipment was 28 degrees Celsius, and the temperature on the suburban equipment was 26.5 degrees Celsius, a difference of only 1.5 degrees Celsius, which did not exceed the limit. ,but The pair is equal to zero. No contribution.

[0134] By introducing an upper limit constraint on the intensity of the urban heat island, a physical consistency constraint is achieved on the model's temperature forecast output. This prevents the model from generating non-physical high-value biases in urban temperature forecasts during training, ensuring that the forecast results conform to the basic laws of urban climate and improving the reliability and operational availability of the forecast results.

[0135] In one embodiment, the training loss function further includes a funnel effect wind speed correction loss term. :

[0136] in A collection of equipment nodes located in densely populated urban high-rise building areas; The wind speed value at this node represents the original weather forecast grid. These are the wind speed forecast values ​​after model correction; This is the correction factor for the aforementioned canyon effect.

[0137] During implementation, the set It includes Substation A within the core business district and six line towers located in a cluster of high-rise buildings to its northwest. The area where these equipment nodes are located has an average building height of over sixty meters and a street width-to-depth ratio greater than 1.5, exhibiting a significant funnel effect.

[0138] Taking substation A as an example, the actual impact of this loss is illustrated. During the outer influence of the typhoon, the output wind speed of the grid points covering the area where substation A is located in the original three-kilometer resolution meteorological forecast grid is... The speed is 12 meters per second. Since the three-kilometer grid is much larger than the spatial scale of a street canyon (approximately 30 to 80 meters), the wind speed within this grid reflects the average wind speed across various terrain and building types, and cannot reflect the amplification effect of local wind speeds within the street canyon. This is as described in claim 9, the canyon effect correction factor for the core business district where substation A is located. Approximately 0.9. The model outputs this value when correcting the wind speed forecast at this equipment location. The value is 16 meters per second, reflecting the model's local correction to the original grid wind speed. Substituting this into the calculation:

[0139] The result was negative. The value is zero, indicating that the device is a point-to-point device. "No contribution" means that although the corrected wind speed of 16 m / s output by the model is higher than the original grid wind speed of 12 m / s, it still does not reach the expected wind speed of 30.4 m / s after correction based on the canyon effect theory. However, according to the design of the loss function, as long as the corrected wind speed is not lower than the original grid wind speed multiplied by the canyon effect correction factor, that is, the model does not correct the wind speed to a lower value than the original, this term will not incur a penalty. This is based on a fundamental judgment: under the condition of limited data, it is a reasonable conservative constraint to prevent the model from negatively underestimating the wind speed in densely populated high-rise areas, thus avoiding predicting a lower value for wind speeds that may actually be higher due to the canyon effect.

[0140] If the model outputs in another forecast The value is 5 meters per second, meaning the model incorrectly corrected the wind speed at that point to be much lower than the original grid value. In this case:

[0141] A value of 2.5 results in a positive loss contribution. Backpropagation will penalize this improper underestimation, prompting adjustments to model parameters to avoid undue suppression of wind speeds for equipment in densely populated high-rise areas.

[0142] By using the narrowing effect wind speed correction loss term, a conservative constraint is achieved on the wind speed forecast of equipment in densely populated high-rise building areas. This ensures that the model output does not systematically underestimate wind speed due to grid resolution limitations in this special urban micro-meteorological environment, thus guaranteeing the reliability of wind load early warning for power grid equipment.

[0143] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A meteorological monitoring method for power grid equipment, characterized in that, include: Obtain a meteorological dataset within the target area. The meteorological dataset contains meteorological monitoring data from multiple sources. The meteorological monitoring data is spatiotemporally aligned using timestamps and coordinate information to obtain spatiotemporally unified multi-source observation data. The multi-grid four-dimensional variational method is used to refine the background field of the multi-source observation data and the numerical model from coarse to fine, and then gradually fused them to obtain a gridded analysis field with a predetermined spatial resolution and a predetermined temporal resolution. The multi-grid four-dimensional variational method is a cost function that includes a background term, an observation term, and a temporal increment smoothing term. A deep residual network is used to extract the low-resolution gridded analysis field and output the corresponding high-resolution grid field. Obtain the geographic coordinate information of the power grid equipment to be mapped within the target area, wherein the power grid equipment includes at least one of transformer substations, power distribution areas, and transmission lines; For each power grid device to be mapped, a set of grid points centered on the geographic coordinates of the power grid device is determined from the high-resolution grid field, and covered by a square window with a preset side length, which is used as the set of grid points in the spatial neighborhood of the power grid device. For each grid point within the neighborhood, the mapping weight of the grid point relative to the power grid equipment is calculated; based on the mapping weight of each grid point within the neighborhood, the meteorological element values ​​of each grid point are normalized and weighted to obtain the accurate meteorological element monitoring values ​​of the location of the power grid equipment.

2. The method according to claim 1, characterized in that, The spatiotemporal alignment is based on the coefficient of variation. An implementation of a cubic spline interpolation algorithm for adaptively adjusting window width, including: The adaptive window cubic spline interpolation algorithm is used to process discrete meteorological monitoring data, wherein the interpolation window width is... From the coefficient of variation of the data ; The adaptive window sliding interpolation algorithm uses cubic spline interpolation, and its interpolation process is expressed as follows: In the formula, These are the interpolated meteorological element values ​​that change continuously over time. For the original set of discrete data points, For timestamps, These are the meteorological element values ​​for the corresponding time. Based on the coefficient of variation Adaptive adjustment of the interpolation window width. Defined as , The standard deviation of the data points within the window. It is the mean, dimensionless.

3. The method according to claim 1, characterized in that, The cost function is: Let be the state vector of the analysis field to be solved at the beginning of the time window, and be the initial condition of the analysis field at all subsequent time points; The state of the background field at the initial moment; The background error covariance matrix describes the spatial correlation between grid points and variables in the background field. The number of data source categories observed; The number of discrete time steps within the time window; For the first Observational data at time The observation vector; For a moment The observation operator maps the analysis field from the model space to the observation point; For a moment The observation error covariance matrix reflects the error characteristics of each observation instrument; For a moment The analytical field state is determined by It is obtained through time integration in a numerical model, i.e. , For model forecasting operators; This is a tangent linear mode, used to linearize the state evolution between adjacent time steps; Let be the mode error covariance matrix, which describes the uncertainty accumulated by the mode during time integration.

4. The method according to claim 1, characterized in that, The deep residual network adopts a deep residual generator structure, extracts multi-scale features through multiple densely connected dilated convolutional blocks, introduces upsampling skip connections to retain large-scale background information, and uses a composite loss function to simultaneously optimize pixel-level mean square error, Laplacian smoothing term and structural similarity index during training.

5. The method according to claim 4, characterized in that, The downscaling mapping relationship of the deep residual network is as follows: in, The output is a high-resolution meteorological field with dimensions of , The number of rows and columns of the high-resolution grid. Number of meteorological element channels; The input is a low-resolution meteorological field with dimension . upsampling factor ; For parameters The deep residual generative network, composed of densely connected dilated convolutional blocks, is used to learn high-frequency detail information; For bicubic interpolation upsampling operators, the output is the same as... Baseline fields of the same dimension.

6. The method according to claim 1, characterized in that, The mapping weights are calculated according to the following formula: in, For power grid equipment points The combined meteorological element values; For equipment points around The set of all high-resolution grid point indices within the window; For grid points Weather values ​​at the location; These are the weight coefficients before normalization; the final weights are obtained through... Divide by Complete normalization; For equipment points With grid points The Euclidean distance between them; This is the distance attenuation scale parameter; The gradient alignment coefficient; Let be the cosine of the angle between the meteorological gradient direction and the direction of the line connecting the grid point to the equipment point, where , For grid points The meteorological gradient vector at that location, To from grid points Pointing to device The vector.

7. The method according to claim 6, characterized in that, The mapping weight calculation also includes an urban canyon effect correction factor. : in, This represents the average height of buildings in the area. The average street width in the area. The depth-to-width ratio of the street canyon reflects the canyon effect; the larger the ratio, the more pronounced the canyon effect. The empirical correction coefficients for different types of built-up areas are determined by referring to a table based on the building height distribution density within the area. For wind speed forecasting using meteorological grids, For reference wind speed; The floor area of ​​the high-rise buildings in this area. The ratio represents the total area of ​​the region. Spatial distribution reflecting regional building density.

8. A method for mapping meteorological forecast data to power grid equipment, characterized in that, include: Using the method described in any one of claims 1-7, accurate meteorological element monitoring values ​​of the location of each power grid device within the target area are obtained; Meteorological numerical forecast data of the target area within a preset forecast period is obtained. The precise meteorological element monitoring values ​​are used as initial state constraints. A spatiotemporal coding network is constructed to extract spatiotemporal features from the meteorological numerical forecast data to obtain forecast spatiotemporal features. Obtain graph structure data of power grid equipment within the target area, wherein the graph structure data includes the geographic coordinate information and static attribute information of each equipment node; An attention mapping network is used to dynamically map the device node as the query and the forecast spatiotemporal features within its spatial neighborhood as the key and value. The attention aggregation result is then fused with the static attribute information of the device node through a gating fusion mechanism to output the refined weather forecast value of the device node.

9. The mapping method according to claim 8, characterized in that, The spatiotemporal coding network is a three-dimensional convolutional spatiotemporal encoder, composed of multiple stacked three-dimensional convolutional layers, and its abstract representation is as follows: in, The input is a numerical weather forecast sequence with dimension [missing information]. ; For time steps, The number of rows and columns of the spatial grid. Number of meteorological element channels; This is a 3D convolutional encoder function, consisting of multiple convolutional layers, batch normalization layers, and ReLU activation functions stacked together; The output is a low-dimensional spatiotemporal feature tensor with dimension . ,in This represents the number of encoder output channels.

10. The mapping method according to claim 8, characterized in that, The attention mapping network is a multi-head cross-attention network, and the gating fusion mechanism uses learnable gating coefficients to weight and combine dynamic features and static attributes. The core operation is represented as follows: in, For device nodes Refined fusion of feature vectors; For nodes The initial forecast feature vector is obtained by bilinear sampling from Based on node coordinates Extracted; For nodes A set of spatially adjacent forecast grid point indices; For neighboring grid points The predicted feature vector; For multi-head cross-attention function, For query, Given keys and values, output the aggregated dynamic features; For device nodes Static attribute encoding vector; It is a gated fusion function that uses learnable gating coefficients to weight and combine dynamic features and static attributes.

Images