High-resolution set numerical weather prediction method, equipment and medium

Abstract

The invention belongs to the technical field of power grids. The high-resolution set numerical weather prediction method comprises the following steps: acquiring geographic information and historical weather information of a target area, and performing grid division on the target area according to the geographic information and the historical weather information to obtain a plurality of grid areas; according to the geographic information and historical meteorological information of the target area, constructing a large high-resolution set numerical weather prediction model, and performing training to obtain a large target weather prediction model; according to the geographic information and historical meteorological information of the target area, generating weather prediction large model parameters of a plurality of grid areas through feature-driven dynamic modeling; using the target weather prediction large model to predict the current weather information of the target area to obtain a prediction result; and adjusting a prediction result according to the weather prediction large model parameters of the plurality of grid regions to obtain a region prediction result of each grid region.

Description

Technical Field

[0001] The present application belongs to the field of power grid technology, and in particular relates to a high-resolution ensemble numerical weather prediction method, device, and medium. Background Art

[0002] With the large-scale integration of distributed photovoltaic power generation into the grid, the inherent randomness and volatility of its power generation output has led to a temporal and spatial mismatch between source and load. This not only exacerbates the difficulty of local photovoltaic consumption but also causes prominent problems such as reverse power flow and voltage limit violations in the distribution network. The current grid operation is characterized by "local problems becoming globalized, and distribution network problems becoming mainstreamed," placing higher demands on the power system's balancing capabilities, reactive power, voltage control, and power quality management. Of particular note, distributed photovoltaic units are small in capacity, dispersed, and numerous. Existing monitoring systems lack comprehensive real-time data collection, making it difficult to timely aggregate operational and forecast data to the dispatching center. This has resulted in a long-term "blind dispatch and control" state for grid dispatching. This data gap directly distorts the load curve. During the critical summer supply guarantee period, inaccurate distributed photovoltaic forecasts directly impact the precise control of the provincial grid's total controlled load. Summary of the Invention

[0003] Based on this, it is necessary to provide a high-resolution ensemble numerical weather prediction method, equipment and medium to address the above technical issues.

[0004] In a first aspect, the present application provides a high-resolution ensemble numerical weather prediction method, comprising: Obtaining geographic information and historical meteorological information of a target area, and dividing the target area into grids based on the geographic information and the historical meteorological information to obtain a plurality of grid areas; Based on the geographic information and historical meteorological information of the target area, a high-resolution ensemble numerical weather forecast model is constructed and trained to obtain a target weather forecast model; Generate weather forecast model parameters for multiple grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area; Using the target weather forecast model, the current weather information of the target area is forecasted to obtain a forecast result; The forecast results are adjusted according to the weather forecast model parameters of the plurality of grid areas to obtain regional forecast results for each of the grid areas.

[0005] In some practicable embodiments, the step of obtaining geographic information and historical meteorological information of a target area, and dividing the target area into grids based on the geographic information and the historical meteorological information to obtain a plurality of grid areas includes: In the target area, taking the area boundary as the smallest unit, grid division of the predicted size is performed to obtain a number of initial grid areas; Performing a first adjustment on a number of initial grid areas according to the cloud image sequence and the irradiance time series in the historical meteorological information to obtain preliminarily adjusted initial grid areas; According to the terrain elevation, slope aspect, and historical sun trajectory in the geographic information, the initially adjusted initial grid area is adjusted for the second time to obtain a grid area.

[0006] In some practicable embodiments, the step of constructing a high-resolution ensemble numerical weather prediction model based on the geographic information and historical meteorological information of the target area, and training the model to obtain the target weather prediction model includes: Constructing the high-resolution ensemble numerical weather forecast model based on the geographic information; Training the high-resolution ensemble numerical weather forecast model based on historical meteorological information to obtain a target weather forecast model; Construct dynamic training sliding windows; Using the dynamic training sliding window, the target weather forecast model is trained in real time to obtain dynamic parameters; The target weather forecast model is updated using the dynamic parameters.

[0007] In some practicable embodiments, the step of generating weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area includes: Extracting features from the geographic information and historical meteorological information of each grid area and fusing feature vectors to obtain a fused feature vector of the grid area; A spatial association graph is constructed using grid areas as nodes and adjacent grid area relationships as edges. Neighborhood features are aggregated based on the neighborhood relationships of each node, and the node state is updated to obtain the final hidden state of the node, wherein the final hidden state represents the geographical and meteorological association features of the node and its neighborhood. According to the final hidden state of the node, it is mapped to physical parameters to form the initial parameters, and the mean and standard deviation of the parameters are calculated based on Monte Carlo Dropout sampling to generate the weather forecast model parameters and their uncertainties, wherein the standard deviation is used to guide the disturbance amplitude of the ensemble forecast.

[0008] In some practicable embodiments, the step of using the target weather forecast model to forecast the current meteorological information of the target area and obtaining the forecast result includes: Extract features from current meteorological information and geographic information, and fuse them to obtain fusion features; The fusion features are input into the target weather forecast model to obtain the forecast results.

[0009] In some practicable embodiments, the step of adjusting the forecast results based on the weather forecast model parameters of the plurality of grid areas to obtain the regional forecast results for each grid area includes: Correcting the forecast result according to the weather forecast model parameters of the grid area to obtain an initial forecast result of the grid area; Monte Carlo perturbation is used to perform parameter perturbation prediction on the initial prediction result of each grid area to obtain a regional prediction result of each grid area.

[0010] In some practicable embodiments, after the step of adjusting the forecast results according to the weather forecast model parameters of the plurality of grid areas to obtain the regional forecast results for each of the grid areas, the method further includes: Calculating the correction parameters of each grid area according to the predicted mean and standard deviation of each grid area and the parameter set generated by Monte Carlo perturbation; The correction parameters are loaded into the target weather forecast model, and the parameters of the target weather forecast model are adjusted to obtain the final target weather forecast model.

[0011] In a second aspect, the present application provides a high-resolution ensemble numerical weather prediction system, which is applied to the aforementioned high-resolution ensemble numerical weather prediction method, and the device includes: an acquisition unit, configured to acquire geographic information and historical meteorological information of a target area, and divide the target area into grids according to the geographic information and the historical meteorological information to obtain a plurality of grid areas; A training unit is used to construct a high-resolution ensemble numerical weather prediction large model based on the geographic information and historical meteorological information of the target area, and to perform training to obtain a target weather prediction large model; A parameter generation unit is used to generate weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area; A prediction result unit is used to use the target weather prediction model to predict the current meteorological information of the target area and obtain a prediction result; The regional prediction result unit is used to adjust the prediction result according to the weather prediction model parameters of multiple grid areas to obtain the regional prediction result of each grid area.

[0012] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the aforementioned high-resolution ensemble numerical weather prediction method when executing the computer program.

[0013] In a fourth aspect, the present application provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned high-resolution ensemble numerical weather prediction method.

[0014] Beneficial Effects: A high-resolution ensemble numerical weather prediction method obtains geographic information and historical meteorological data for a target area and divides the target area into multiple grid regions based on this information. A high-resolution ensemble numerical weather prediction model is constructed and trained based on the target area's geographic information and historical meteorological data to obtain the target weather prediction model. Based on the target area's geographic information and historical meteorological data, feature-driven dynamic modeling is used to generate weather prediction model parameters for the multiple grid regions. The target weather prediction model is used to predict the current weather information in the target area, obtaining prediction results. The prediction results are then adjusted based on the weather prediction model parameters for the multiple grid regions to obtain regional prediction results for each grid region. This method, based on the target area gridding technique based on geographic information and historical meteorological data, enables a refined characterization of meteorological conditions in the region where the photovoltaic cluster is located. Furthermore, the constructed high-resolution ensemble numerical weather prediction model establishes a meteorological prediction framework that accounts for local terrain effects. A prediction output containing uncertainty information is provided for each grid region. This ensemble prediction approach not only provides deterministic forecast values but also quantifies the reliability of the forecast results, providing critical meteorological input support for photovoltaic power forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 The present invention is a flowchart of a high-resolution ensemble numerical weather prediction method in one embodiment. DETAILED DESCRIPTION

[0017] To facilitate understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The accompanying drawings provide embodiments of the present application. However, the present application may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The term "and / or" as used herein includes any and all couplings of one or more of the associated listed items.

[0019] It will be understood that the terms "first," "second," etc. used herein may be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish a first element from another element.

[0020] The following are some explanations of some terms involved in this application to facilitate understanding of this application: In the power grid GIS (Geographic Information System), topology adjustment of the grid (such as cutting or merging) refers to the dynamic division or reorganization of the physical or logical structure of the power grid to adapt to operational needs, optimize management or improve analysis accuracy; grid cutting is to split a grid in the original power grid topology (such as a substation, feeder, or distribution area) into multiple smaller grid units; grid merging is to merge multiple adjacent grids into a larger grid unit.

[0021] Centroid coordinate calculation is used to determine the geometric center or electrical load center of the power grid (such as the substation area or feeder area), assisting in grid cutting, merging, or resource optimization deployment.

[0022] ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) is a multispectral remote sensing sensor. It provides high-resolution topography, surface temperature, and reflectivity data, which can be used for terrain modeling, line planning, disaster assessment, and other scenarios in power grid GIS topology analysis.

[0023] WRF (Weather Research and Forecasting Model) is a mesoscale numerical weather forecast system that is widely used in meteorology, climate, energy (wind power / photovoltaic forecasting), disaster warning and other fields.

[0024] Graph Convolutional Networks (GCN) extend traditional convolution operations to graph-structured data (non-Euclidean space) and generate high-order node representations by aggregating the node's own features and its neighbors' features.

[0025] like Figure 1 As shown, the present application provides a high-resolution ensemble numerical weather prediction method, comprising: S100, obtaining geographic information and historical meteorological information of a target area, and dividing the target area into grids according to the geographic information and the historical meteorological information to obtain a plurality of grid areas.

[0026] Specifically, obtaining multiple grid areas may include the following steps: S101 , in the target area, taking the area boundary as the smallest unit, performing grid division of a predicted size to obtain a number of initial grid areas.

[0027] Using the grid boundary as the smallest unit, a grid covering all PV clusters is generated, such as a 1 km x 1 km grid. If the grid area is less than 1 km², it is forcibly merged with adjacent grids to form a joint grid. If the grid spans multiple 1 km grids, it is split based on "center point ownership." For example, center point ownership can be calculated using centroid coordinates. This method generates a basic grid area, which can be adjusted in subsequent steps using historical meteorological and geographic information.

[0028] S102 , performing a first adjustment on a number of initial grid areas according to the cloud image sequence and the irradiance time series in the historical meteorological information to obtain preliminarily adjusted initial grid areas.

[0029] Specifically, the cloud image sequence includes the cloud movement velocity field and the cloud cover time probability.

[0030] By quantitatively analyzing the cloud motion characteristics and the spatial variation of irradiance, grid optimization driven by meteorological characteristics is implemented.

[0031] The calculation of the cloud movement velocity field is based on the time-series motion vector data obtained by the optical flow analysis method. This method tracks the displacement of feature points in continuous cloud images and uses the principle of pixel brightness conservation to invert the cloud movement trajectory. The calculation formula for the cloud movement velocity field is: ; Corresponding to the x and y velocity components of the optical flow field at time step t (unit: m / s), T represents the total number of time steps of the historical cloud map.

[0032] Optical flow analysis is used to process a continuous sequence of historical satellite cloud images, extracting the cloud motion vector field at each time step t. By analyzing the displacement of cloud feature points (such as cloud top temperature gradients and texture characteristics) in adjacent cloud images, the velocity components in the x and y directions are obtained for each grid point (x, y). The arithmetic mean of the instantaneous velocity modulus across all time steps is taken to obtain the average cloud motion velocity at the grid point (x, y).

[0033] The formula is used to calculate the average cloud movement speed of the grid point (x,y) in the time series. Next, calculate the speed standard deviation : ; in, represents the cloud speed scalar (modulus) at time step t; , represents the average cloud moving speed; Speed standard deviation , reflects the degree of fluctuation of cloud movement speed. The larger the value, the more unstable the speed.

[0034] Cloud frequency indicates the percentage of time the grid is obscured by clouds, reflecting the risk of power generation loss.

[0035] Cloud frequency ( )formula: : Number of cloud cover time steps (i.e., time steps with cloud cover > 50%).

[0036] Determine encryption conditions: If >60%, the grid needs to be refined.

[0037] Cloud cover time probability : Represents the cloud obscuration probability (range 0-1) at the grid point (x,y).

[0038] Determine encryption conditions: >0.6, which means that the area is obscured by clouds more than 60% of the time.

[0039] For example, the conditions for refining the initial mesh area (increasing the mesh area density) are: Speed standard deviation: the degree of fluctuation in cloud movement speed (standard deviation > 2 m / s); Cloud cover time probability: the proportion of time when cloud cover occurs (> 60%).

[0040] Irradiance time series refers to the irradiance observations of a single grid at different time points, which is used to count the historical frequency of LSVI (Neighborhood Irradiance Spatial Variability Index) exceeding the standard. Within a grid area, the LSVI of different areas is: n×n grid centered at (x,y) (e.g., covering a 3 km×3 km area); Determine encryption conditions: >0.25 indicates that the irradiance fluctuates violently in the local space (such as sudden changes caused by the rapid movement of cumulus clouds).

[0041] For example, the encryption operation is performed: Encryption conditions (encryption occurs if any of the conditions are met): 1. The standard deviation of cloud movement speed is > 2 m / s; 2. Cloud cover probability > 60%; 3. Irradiance LSVI > 0.25; Encryption method: Divide the 1 km × 1 km grid into four equal parts of 0.5 km × 0.5 km sub-grids.

[0042] S103 , performing a second adjustment on the preliminarily adjusted initial grid area according to the terrain elevation, slope aspect, and historical sun trajectory in the geographic information to obtain a grid area.

[0043] Specifically, the actual sunshine duration is calculated through DEM (terrain elevation), taking terrain shading into account.

[0044] Historical sun trajectory: provides theoretical maximum sunshine duration (such as historical data of solar altitude and azimuth angles).

[0045] The sunshine duration weight is calculated based on the actual sunshine duration and the theoretical maximum sunshine duration.

[0046] Exemplarily, the encryption condition is judged to have a sunshine duration weight greater than 0.3.

[0047] Aspect-irradiance matching (aspect): Aspect indicates the direction of the surface slope, starting from true north (defined as 0°) and rotating clockwise to 360°, covering all directions. Aspect is the grid-averaged aspect (Aspect(x,y)) calculated based on the DEM. For example, when the aspect is 180° (due south), the deviation angle is 0°, indicating a complete south orientation. When the aspect is 0° (due north) or 360°, the deviation angle is 180°, indicating a direction completely opposite to due south.

[0048] For example, the conditions for judging the encryption are: the slope is greater than 45°, that is, the photovoltaic efficiency of areas where the slope deviates from due south by more than 45° (such as north-facing slopes) is low, and encryption is required to improve the prediction accuracy.

[0049] For example, the dynamic adjustment rule (terrain + solar synergy), i.e., secondary encryption, is performed as follows: Condition: For grids that have been densified to 0.5 km, if the sunshine duration weight is greater than 0.3 (terrain shadow) and the slope aspect is less than 45° (slope mismatch), further densification to 0.25 km is allowed; Operation: Densify to 0.25 km while responding to the optimization needs of terrain elevation and slope.

[0050] Mesh merging rules: Condition: 1 km grids with sunshine duration weight less than 0.1 (low shadow) and no photovoltaic installations are merged into 2 km grids to reduce the computational effort; Action: Merge to 2 km and reduce the computational effort in non-critical areas to retain the necessary accuracy for terrain-sun coupling.

[0051] It should be noted that each substation (distribution unit) contains at least one complete grid to avoid cross-substation splitting. If the substation boundary does not match the grid, the grid should be adjusted (such as cutting or merging) according to the power grid GIS topology.

[0052] S200: construct a high-resolution ensemble numerical weather forecast model based on the geographic information and historical meteorological information of the target area, and perform training to obtain a target weather forecast model.

[0053] Specifically, obtaining the target weather forecast model may include the following steps: S201, constructing the high-resolution ensemble numerical weather forecasting model according to the geographic information.

[0054] Specifically, the WRF model uses known high-resolution terrain data (such as ASTER, with a resolution of ≥30m) to optimize the terrain input to accurately match the target area's landforms, such as mountains and rivers. In other words, using known geographic information, the target area is divided into 1km×1km grids, ensuring that the terrain parameters (altitude, slope, and surface type) of each grid are consistent with the actual terrain.

[0055] S202: Training the high-resolution ensemble numerical weather prediction model based on historical meteorological information to obtain a target weather prediction model.

[0056] Specifically, the high-resolution ensemble numerical weather forecast model is trained using historical meteorological information for the target region, such as the past three years. This means that before training, the high-resolution ensemble numerical weather forecast model is an initial, conventional model. Training this model with historical meteorological information enables it to predict the historical meteorological information for the target region. Once the high-resolution ensemble numerical weather forecast model is trained with historical meteorological information, the target weather forecast model is obtained.

[0057] The historical meteorological information has been mentioned in the relevant steps of S100 and will not be described again here.

[0058] After obtaining the target weather forecast model, the meteorological information of the current window can be used as input to perform forecasting using the target weather forecast model.

[0059] S203: Build a dynamic training sliding window.

[0060] For example, a window rule might set an initial training window length of 30 days, with a sliding window updated daily. Furthermore, the window length can be dynamically adjusted based on the stability of historical meteorological data (e.g., the rate of change of variance), with a minimum of 7 days and a maximum of 60 days. Specifically, the window length is adjusted based on the rate of change of variance, a threshold is set (e.g., 1.5 times the historical average error variance), and the variance of the window length is compared with the threshold. For example, if the variance exceeds the threshold (e.g., 1.5 times the historical average error variance), the window length is shortened to 7 days to quickly respond to sudden weather changes; if the variance is below the threshold, the window length is extended to 60 days to maintain stability.

[0061] S204: Using the dynamic training sliding window, perform real-time training on the target weather forecast model to obtain dynamic parameters.

[0062] Specifically, the meteorological information within the dynamic training sliding window is used as input data into the target weather forecast model for real-time training. This allows for timely parameter adjustments to ensure the model's timeliness. In other words, dynamic training of the sliding window yields dynamic parameters for the target weather forecast model. For example, Bayesian model averaging (BMA) is employed to calculate the target weather forecast model's weight coefficients and error correction parameters based on the data within the window. The essence of dynamic training of the sliding window is to utilize meteorological information within the window period to promptly update the parameters of the target weather forecast model, ensuring the timeliness and accuracy of the model's predictions.

[0063] S205: Utilize the dynamic parameters to update the target weather forecast model.

[0064] Specifically, the dynamic parameters are fed back to the target weather forecast model to update the parameters of the target weather forecast model.

[0065] S300, generating weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area.

[0066] Specifically, generating weather forecast model parameters for a plurality of grid areas may include the following steps: S301 , performing feature extraction on the geographic information and historical meteorological information of each grid area, and performing feature vector fusion to obtain a fused feature vector of the grid area.

[0067] Specifically, the contents of geographic information and historical meteorological information can be adjusted according to actual conditions. For example, geographic information includes geographic features such as elevation, slope, and surface type coding; meteorological information may include meteorological features such as cloud obscuration probability, spatial variation index of irradiance, and standard deviation of cloud movement speed.

[0068] Regular standardization is performed on geographic features and meteorological features. After standardization, feature vector fusion is performed to obtain the fused feature vector of the grid area, so as to solve the problem of heterogeneous data fusion of geographic features and meteorological features.

[0069] For example, a normalization process is first performed: the value of each feature is first subtracted from the mean value of that feature across all grids, and then divided by its standard deviation. This process eliminates the dimensional differences between different features, making heterogeneous data such as mountain elevation (unit: meters) and cloud cover probability (0-1 scale) comparable.

[0070] Next, dynamic weighting of the fused features is performed. The historical volatility (variance) of meteorological features is calculated, with more volatile features having a greater impact on the forecast. A temperature-controlled weighting mechanism increases the weight of high-variance features (such as sudden changes in irradiance during typhoons) to 60%-80%, while the weight of stable features (such as terrain slope) is correspondingly reduced to 10%-20%. This dynamic adjustment enables the model to adaptively prioritize and capture sudden meteorological events.

[0071] Finally, the standardized geographic and meteorological features are weighted and concatenated according to their assigned weights. For example, the processed elevation (weight 0.15) and slope (weight 0.18) of a given grid are combined with the real-time cloud cover probability (weight 0.42) and irradiance variation (weight 0.25) to generate a four-dimensional feature vector. This fusion preserves the inherent properties of the terrain while highlighting dynamic meteorological changes, resulting in a 32-dimensional comprehensive feature representation (the feature dimension is higher in actual engineering).

[0072] S302, constructing a spatial association graph with grid areas as nodes and adjacent grid area relationships as edges, and performing neighborhood feature aggregation based on the neighborhood relationship of each node, updating the state of the node, and obtaining the final hidden state of the node.

[0073] The final hidden state represents the geographical and meteorological correlation characteristics of the node and its neighborhood.

[0074] Specifically, the fused feature vectors of all grids and each grid are connected to its 8-neighboring grids (up, down, left, right, and diagonal) to complete the graph structure construction.

[0075] Graph structure construction: Node: Each geographic grid cell is mapped to an independent node in the graph structure. Each node carries a standardized fused feature vector (including geographic attributes such as elevation and slope, and meteorological characteristics such as cloud cover probability and irradiance variation). Edges: Based on the spatial adjacency of grids, eight-neighborhood connections are established. If two grids are spatially adjacent (including vertically, horizontally, and diagonally), undirected edges are created, forming a network of connections covering the entire area. For example, the central grid will be connected to the eight surrounding grids, reflecting the continuity of the terrain and the spatial propagation characteristics of meteorological elements.

[0076] For example, neighborhood feature aggregation can be achieved through a two-layer graph convolutional network (GCN) to achieve feature propagation and fusion. The specific process is as follows: The first convolutional layer (local feature extraction) takes as input the initial feature vector (dimension 4) of each node. Neighborhood aggregation aggregates the feature information of each node's immediate neighbors (eight neighbors). This calculation involves a normalized weighted sum of the neighboring features, with the weight determined by the inverse square root of the node's degree. A nonlinear transformation linearly transforms the aggregation result using a learnable weight matrix (dimension 4×16) to produce a 16-dimensional intermediate feature vector that captures local terrain-meteorological correlations (such as the cloud accumulation effect caused by mountain obstruction).

[0077] The second convolutional layer (global correlation modeling) takes the 16-dimensional feature vector output by the first layer and propagates it across layers, further aggregating information from the extended neighborhood (indirect neighbors). The weight matrix (16×32 dimensions) learns high-order spatial patterns. It outputs a final 32-dimensional hidden state, encoding the geographic and meteorological coupling between the node itself and its multi-hop neighborhood (approximately 64 grid cells), such as the guiding effect of mountain range orientation on cloud migration. The numbers for the first and second layers are for illustrative purposes only.

[0078] The significance of the final hidden state lies in the fact that the 32-dimensional hidden state vector of each node comprehensively represents the following related features: terrain modulation effect, the influence of elevation gradient on cloud lift / subsidence; meteorological propagation path, the spatial synergistic relationship between cloud speed standard deviation and slope aspect; irradiance mutation chain, the transmission pattern of LSVI high value areas between adjacent grids.

[0079] The effect of the final hidden state is that if it is passed through two layers of GCN, it can effectively capture the terrain-meteorological interaction within a predetermined range; the eight-neighborhood connection reduces the number of parameters by 98% compared to the fully connected network and increases the training speed by 15 times; the hidden state vector can identify high-risk meteorological patterns through cluster analysis (such as "valley cloud vortex" corresponding to a specific 32-dimensional vector pattern).

[0080] S303, mapping the final hidden state of the node to physical parameters to form initial parameters, and calculating the mean and standard deviation of the parameters based on Monte Carlo Dropout sampling to generate weather forecast model parameters and their uncertainties.

[0081] The standard deviation is used to guide the disturbance amplitude of the ensemble forecast.

[0082] Specifically, the node hidden state vector (e.g., 32-dimensional) obtained in the previous step is mapped to physical parameters. That is, through a trainable fully connected layer, the 32-dimensional hidden state vector is mapped to a set of physical parameters. This may include a weight matrix to learn the correlation between hidden state features and meteorological quantities (such as boundary layer height and surface albedo); a bias term to compensate for regional climate background values (such as base temperature offsets at high altitudes); and output dimensions, set according to forecast requirements, i.e., initial parameters, for example, generating two key parameters (turbulence coefficient and cloud water content).

[0083] Next, after obtaining the settings based on the forecast requirements, such as two key parameters, Monte Carlo Dropout is used to generate a parameter set through independent forward propagation. For example, through 50 independent forward propagations, some neurons are randomly blocked during each propagation, forcing the model to make predictions through different sub-networks. Next, the results are statistically calculated, which may include homogeneity and standard deviation. The homogeneity can be the average of the 50 prediction results as the final physical parameter estimate; the standard deviation can measure the parameter fluctuation range. The larger the value, the higher the model's cognitive uncertainty in the area.

[0084] S400: Utilize the target weather forecast model to forecast the current weather information of the target area and obtain a forecast result.

[0085] Specifically, obtaining the prediction result may include the following steps: S401, extracting features from current meteorological information and geographic information, and fusing them to obtain fused features; S402: Input the fusion features into a target weather forecast model to obtain a forecast result.

[0086] Specifically, existing meteorological information collection equipment is used to collect current meteorological information, and the collected meteorological information is fused with the features of geographic information to obtain fused features. Next, the fused features are input into the trained target weather prediction model. Finally, the target weather prediction model is used to output the prediction results, which represent the weather conditions in the target area.

[0087] It should be noted that feature extraction and fusion are conventional methods, such as using GCN and attention mechanism.

[0088] S500, adjusting the prediction results according to the weather prediction model parameters of the plurality of grid areas to obtain a regional prediction result for each of the grid areas.

[0089] Specifically, obtaining the regional prediction result for each grid area may include the following steps: S501, correcting the forecast result according to the weather forecast model parameters of the grid area to obtain an initial forecast result of the grid area.

[0090] It should be noted that, based on the weather forecast model parameters of the grid area, the statistical characteristics of each parameter are obtained, and targeted adjustments are made to the original numerical forecast results to resolve the systematic deviations caused by the coupling of local terrain and micrometeorology.

[0091] For example, in the correction of radiation prediction in desert areas, the target weather prediction model of a photovoltaic power station in the Taklimakan Desert in Xinjiang predicted a midday irradiance of 1120 W / m², but did not consider the scattering effect of suspended sand and dust particles.

[0092] Among them, the irradiance of 1120 W / m² is formed, and the reference parameters may include aerosol optical depth and surface albedo. Therefore, the two parameters are corrected using the weather prediction model parameters of the grid area to obtain the output prediction results that are consistent with the grid area.

[0093] S502 , performing parameter perturbation prediction on the initial prediction result of each of the grid areas using Monte Carlo perturbation to obtain a regional prediction result for each of the grid areas.

[0094] It should be noted that a Monte Carlo perturbation is constructed (the construction method is conventional), and the Monte Carlo perturbation is applied to the initial prediction result to obtain a regional prediction result for each of the grid areas.

[0095] It should be noted that, for example, 50 predictions are performed for a grid area to form an ensemble prediction, which is then statistically analyzed to form a prediction mean and a prediction standard deviation. The prediction mean and prediction standard deviation are then used as the initial regional prediction result for each grid area (the prediction mean and prediction standard deviation are summed).

[0096] After step S500, the following steps may be further included: S601 , calculating correction parameters for each of the grid areas according to the predicted mean and standard deviation of each of the grid areas and a parameter set generated by Monte Carlo perturbation.

[0097] Specifically, in the previous steps, the mean and standard deviation, as well as the parameter set generated by Monte Carlo perturbation, have been obtained. Next, similar features are grid clustered, retaining major differences (mountains / plains) to meet high-resolution requirements. Among them, the weighted average of similar grid parameters is used. In addition, since the older the parameter, the smaller its impact, the attenuation coefficient of the historical parameters is added to the fused parameters to balance the influence of new and old parameters and prevent parameter oscillation.

[0098] For example, in a seven-day period, daily parameters within the seven-day period are first weighted downwards from near to far, and a weighted average is calculated. Next, the resulting weights are re-weighted, and the corresponding parameters of the target weather forecast model are also weighted. However, the target weather forecast model's weights are smaller than those for the entire period. This double redistribution of weights ensures the accuracy and real-time performance of the target weather forecast model. It should be noted that the weight setting can be adjusted as needed, such as using a dynamic weight adjustment strategy or a fixed weighting strategy. The dynamic weight adjustment strategy uses real-time assessment of the uncertainty (standard deviation) of model parameters and historical data to adaptively assign weights to the two, with data sources with lower uncertainty receiving higher weights. A time decay mechanism is also introduced to exponentially reduce the contribution of historical parameters, ensuring that new data dominates the correction process. This dual weighting mechanism suppresses parameter fluctuations while balancing model stability and real-time performance requirements.

[0099] S602, loading the correction parameters into the target weather forecast model, adjusting the parameters of the target weather forecast model, and obtaining a final target weather forecast model.

[0100] The modified parameters are loaded into the target weather forecast model to obtain the final target weather forecast model. Next, the final target weather forecast model is used to conduct the next round of forecasts.

[0101] In a second aspect, the present application provides a high-resolution ensemble numerical weather prediction system, which is applied to the aforementioned high-resolution ensemble numerical weather prediction method, and the device includes: an acquisition unit, configured to acquire geographic information and historical meteorological information of a target area, and divide the target area into grids according to the geographic information and the historical meteorological information to obtain a plurality of grid areas; A training unit is used to construct a high-resolution ensemble numerical weather prediction large model based on the geographic information and historical meteorological information of the target area, and to perform training to obtain a target weather prediction large model; A parameter generation unit is used to generate weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area; A prediction result unit is used to use the target weather prediction model to predict the current meteorological information of the target area and obtain a prediction result; The regional prediction result unit is used to adjust the prediction result according to the weather prediction model parameters of multiple grid areas to obtain the regional prediction result of each grid area.

[0102] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the aforementioned high-resolution ensemble numerical weather prediction method when executing the computer program.

[0103] In a fourth aspect, the present application provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned high-resolution ensemble numerical weather prediction method.

[0104] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0105] The various embodiments in the present disclosure are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0106] The scope of protection of the present disclosure is not limited to the above-described embodiments. Obviously, those skilled in the art may make various modifications and variations to the present disclosure without departing from the scope and spirit of the present disclosure. If such modifications and variations fall within the scope of the claims of the present disclosure and their equivalents, the present disclosure is intended to include such modifications and variations.

Claims

1. A high-resolution ensemble numerical weather forecasting method, characterized in that: Methods include: Obtaining geographic information and historical meteorological information of a target area, and dividing the target area into grids based on the geographic information and the historical meteorological information to obtain a plurality of grid areas; Based on the geographic information and historical meteorological information of the target area, a high-resolution ensemble numerical weather forecast model is constructed and trained to obtain a target weather forecast model; Generate weather forecast model parameters for multiple grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area; Using the target weather forecast model, the current weather information of the target area is forecasted to obtain a forecast result; The forecast results are adjusted according to the weather forecast model parameters of the plurality of grid areas to obtain regional forecast results for each of the grid areas.

2. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: The step of obtaining geographic information and historical meteorological information of a target area, and dividing the target area into grids based on the geographic information and the historical meteorological information to obtain a plurality of grid areas, includes: In the target area, taking the area boundary as the smallest unit, grid division of the predicted size is performed to obtain a number of initial grid areas; Performing a first adjustment on a number of initial grid areas according to the cloud image sequence and the irradiance time series in the historical meteorological information to obtain preliminarily adjusted initial grid areas; According to the terrain elevation, slope aspect, and historical sun trajectory in the geographic information, the initially adjusted initial grid area is adjusted for the second time to obtain a grid area.

3. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: The step of constructing a high-resolution ensemble numerical weather forecast model based on the geographic information and historical meteorological information of the target area and training the model to obtain the target weather forecast model includes: Constructing the high-resolution ensemble numerical weather forecast model based on the geographic information; Training the high-resolution ensemble numerical weather forecast model based on historical meteorological information to obtain a target weather forecast model; Construct dynamic training sliding windows; Using the dynamic training sliding window, the target weather forecast model is trained in real time to obtain dynamic parameters; The target weather forecast model is updated using the dynamic parameters.

4. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: The step of generating weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area comprises: Extracting features from the geographic information and historical meteorological information of each grid area and fusing feature vectors to obtain a fused feature vector of the grid area; A spatial association graph is constructed using grid areas as nodes and adjacent grid area relationships as edges. Neighborhood features are aggregated based on the neighborhood relationships of each node, and the node state is updated to obtain the final hidden state of the node, wherein the final hidden state represents the geographical and meteorological association features of the node and its neighborhood. According to the final hidden state of the node, it is mapped to physical parameters to form the initial parameters, and the mean and standard deviation of the parameters are calculated based on Monte Carlo Dropout sampling to generate the weather forecast model parameters and their uncertainties, wherein the standard deviation is used to guide the disturbance amplitude of the ensemble forecast.

5. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: The step of using the target weather forecast model to forecast the current meteorological information of the target area to obtain a forecast result includes: Extract features from current meteorological information and geographic information, and fuse them to obtain fusion features; The fusion features are input into the target weather forecast model to obtain the forecast results.

6. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: The step of adjusting the prediction results according to the weather prediction model parameters of the plurality of grid areas to obtain the regional prediction results for each of the grid areas comprises: Correcting the forecast result according to the weather forecast model parameters of the grid area to obtain an initial forecast result of the grid area; Monte Carlo perturbation is used to perform parameter perturbation prediction on the initial prediction result of each grid area to obtain a regional prediction result of each grid area.

7. The high-resolution ensemble numerical weather forecasting method according to claim 1, characterized in that: After the step of adjusting the forecast results according to the weather forecast model parameters of the plurality of grid areas to obtain the regional forecast results for each of the grid areas, the method further includes: Calculating the correction parameters of each grid area according to the predicted mean and standard deviation of each grid area and the parameter set generated by Monte Carlo perturbation; The correction parameters are loaded into the target weather forecast model, and the parameters of the target weather forecast model are adjusted to obtain the final target weather forecast model.

8. A high-resolution ensemble numerical weather prediction system, characterized by: The high-resolution ensemble numerical weather prediction method according to any one of claims 1 to 7, wherein the device comprises: an acquisition unit, configured to acquire geographic information and historical meteorological information of a target area, and divide the target area into grids according to the geographic information and the historical meteorological information to obtain a plurality of grid areas; A training unit is used to construct a high-resolution ensemble numerical weather prediction large model based on the geographic information and historical meteorological information of the target area, and to perform training to obtain a target weather prediction large model; A parameter generation unit is used to generate weather forecast model parameters for a plurality of grid areas through feature-driven dynamic modeling based on the geographic information and historical meteorological information of the target area; A prediction result unit is used to use the target weather prediction model to predict the current meteorological information of the target area and obtain a prediction result; The regional prediction result unit is used to adjust the prediction result according to the weather prediction model parameters of multiple grid areas to obtain the regional prediction result of each grid area.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the high-resolution ensemble numerical weather prediction method according to any one of claims 1 to 7 are implemented.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the high-resolution ensemble numerical weather prediction method according to any one of claims 1 to 7 are implemented.

Images