A regional weather prediction method and system based on a meteorological large model

Abstract

The application relates to the technical field of policy optimization, and discloses a regional weather prediction method and system based on a meteorological large model, which comprises the following steps: acquiring a terrain influence weight matrix; inputting the weight matrix into a spatial distribution feature model to output an optimized spatial distribution feature set; performing spatio-temporal registration through Kriging interpolation to obtain a spatial distribution feature set of local weather; extracting a seasonal change trend, performing deviation correction, and obtaining a deviation correction coefficient; updating the weight matrix according to the deviation correction coefficient and performing coupling calculation to generate a coupling parameter set; generating a prediction result according to the coupling parameter set; if the accuracy is lower than a threshold value, generating a calculation resource configuration; combining the resource configuration, the updated weight matrix and the spatial feature set to perform weather prediction and obtain a prediction result; and finally generating a rainfall intensity distribution map and performing differential correction to output a regional weather prediction result. The application can realize high-precision prediction of complex regional weather.

Description

Technical Field

[0001] This invention relates to the field of strategy optimization technology, and in particular to a regional weather forecasting method and system based on a large meteorological model. Background Technology

[0002] Weather forecasting, as a crucial component of the public service system, plays a fundamental supporting role in social operation and resource allocation. It not only provides a basis for government departments to formulate disaster prevention and mitigation plans, but also offers important references for agricultural production arrangements, transportation organization, and energy dispatch. At the same time, accurate weather forecasts help the public rationally plan their travel and daily lives, reduce the adverse effects of extreme weather, and enhance the overall social response capacity and safety level.

[0003] Current weather forecasting relies on numerical weather prediction (NWP) as its core method. This method establishes atmospheric dynamics and thermodynamic equations and performs numerical integration using real-time observational data to simulate and predict future meteorological elements. While this approach can provide relatively systematic and comprehensive forecasts on medium- to long-term scales, the accuracy of NWP in regions with complex terrain or variable climates is often limited by model resolution, initial field accuracy, and the adaptability of physical parameterization schemes, making it difficult to accurately characterize local weather variations. For example, in mountainous areas, the disturbance of airflow by complex terrain often leads to biases in local precipitation forecasts; in coastal areas, the dynamic changes of sea-land breeze effects are also difficult to fully capture. These shortcomings directly affect the accuracy and reliability of the forecast results.

[0004] In summary, existing technologies suffer from insufficient prediction accuracy in areas with complex terrain or variable climate. Summary of the Invention

[0005] This invention provides a regional weather forecasting method and system based on a large meteorological model to achieve high-precision weather forecasting in complex regions.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a regional weather forecasting method based on a large meteorological model, comprising:

[0007] Obtain the terrain influence weight matrix;

[0008] The terrain influence weight matrix is ​​input into a pre-trained spatial distribution feature model, which outputs an optimized spatial distribution feature set. Spatiotemporal registration is then performed to obtain the spatial distribution feature set of local weather.

[0009] Based on the spatial distribution feature set, seasonal variation trends are extracted to obtain the seasonal variation trends;

[0010] Based on the aforementioned seasonal variation trend, seasonal deviation correction is performed to obtain the deviation correction coefficient;

[0011] Based on the deviation correction coefficient, the terrain influence weight matrix is ​​updated to obtain the updated terrain influence weight matrix, and the terrain influence weight matrix is ​​coupled to obtain the coupling parameter set of deep coupling of regional characteristics.

[0012] Based on the set of coupling parameters, a prediction result is generated. If the prediction result is lower than a preset accuracy threshold, a computing resource configuration is generated to obtain the computing resource configuration.

[0013] Based on the aforementioned computing resource allocation, the updated terrain influence weight matrix, and the aforementioned spatial distribution feature set, weather forecasting is performed to obtain regional weather forecast results.

[0014] Based on the regional weather forecast results, a spatial distribution map is generated to obtain a spatial distribution map of local rainfall intensity, and differential corrections are performed to obtain the regional weather forecast output.

[0015] Preferably, the step of inputting the terrain influence weight matrix into a pre-trained spatial distribution feature model, outputting an optimized spatial distribution feature set, and registering it with spatiotemporal grid data to obtain a spatial distribution feature set of local weather includes:

[0016] Acquire time grid data;

[0017] The spatial distribution feature model is obtained by training a CNN neural network;

[0018] Through the input layer of the spatial distribution feature model, the terrain influence weight matrix is ​​normalized and spatially encoded to obtain a standard spatial weight tensor.

[0019] Through the convolutional layers of the spatial distribution feature model, multi-layer convolution operations are performed on the standard spatial weight tensor to extract local gradient and slope aspect features, thereby obtaining a preliminary spatial feature set.

[0020] The initial spatial feature set is subjected to max pooling dimensionality reduction through the pooling layer of the spatial distribution feature model to obtain the dimensionality-reduced spatial feature set.

[0021] The reduced spatial feature set is integrated through the fully connected layer of the spatial distribution feature model, and an optimized spatial distribution feature set is generated by nonlinear mapping through an activation function.

[0022] Based on the Kriging interpolation method and combined with the time grid data, missing values ​​are filled in the optimized spatial distribution feature set to obtain the spatial distribution feature set of local weather.

[0023] Preferably, the step of extracting seasonal variation trends based on the spatial distribution feature set to obtain seasonal variation trends includes:

[0024] Based on the spatial distribution feature set, Fourier transform is used to extract the main frequency component and calculate the main frequency energy ratio;

[0025] If the proportion of the dominant frequency energy is higher than the preset seasonal variability threshold, the spatial distribution feature set is input into the pre-trained time series model and a preliminary time series feature set is output.

[0026] The time series model is trained using an LSTM neural network;

[0027] Through the input layer of the time series model, the spatial distribution feature set is standardized and time series encoded to generate a standard time series tensor.

[0028] By using the hidden layer of the time series model, state memory features are extracted from the standard time series tensor to obtain the state memory feature sequence;

[0029] Through the output layer of the time series model, sequence mapping and feature optimization are performed on the state memory feature sequence to output a preliminary time series feature set;

[0030] The initial time series feature set is subjected to dimensionality reduction processing, and the main time series features are retained by principal component analysis to obtain the dimensionality-reduced time series feature set.

[0031] The reduced-dimensional time series feature set is subjected to nonlinear mapping to generate optimized time series features;

[0032] Based on the Kriging interpolation method and combined with the time grid data, missing values ​​are filled in for the optimized time series features to obtain the seasonal variation trend.

[0033] Preferably, the step of performing seasonal deviation correction based on the seasonal variation trend to obtain the deviation correction coefficient includes:

[0034] Acquire real-time environmental data and seasonal feature databases;

[0035] Based on the mean filtering method, the real-time environmental data is smoothed to obtain a smoothed environmental data set;

[0036] The difference between the environmental data set and the seasonal feature database is calculated to obtain the deviation value. The deviation is then corrected by combining the seasonal change trend to obtain an initial deviation value set.

[0037] If the absolute value of any parameter in the initial deviation value set exceeds the corresponding preset deviation threshold, the offset parameter is adjusted to obtain the adjusted offset parameter set, and an optimized deviation correction coefficient is generated through a linear transformation method.

[0038] Preferably, the step of updating the terrain influence weight matrix based on the deviation correction coefficient to obtain an updated terrain influence weight matrix, and generating a parameter set to obtain a coupling parameter set for deep coupling of regional characteristics, includes:

[0039] Acquire regional prior data and regional characteristic data;

[0040] Based on the deviation correction coefficient, the terrain influence weight matrix is ​​weighted and adjusted to obtain the updated initial weight matrix;

[0041] Based on the initial weight matrix and the prior data of the region, local sensitivity factors are generated to obtain a set of local sensitivity factors.

[0042] If the set of local sensitive factors meets the preset sensitivity judgment conditions, the initial weight matrix is ​​gradually adjusted to obtain the updated terrain influence weight matrix, and combined with the regional characteristic data, deep coupling calculation is performed to obtain the coupling parameter set of deep coupling of regional characteristics.

[0043] Preferably, the step of generating prediction results based on the coupling parameter set, and generating computing resource configuration if the prediction results are lower than a preset accuracy threshold, includes:

[0044] Acquire performance index data and measured meteorological data for the target area;

[0045] Based on the set of coupling parameters and combined with the regional characteristic data, weather forecasting is performed to obtain preliminary forecast meteorological data for the target area.

[0046] The prediction error is calculated by comparing the preliminary meteorological data with the measured meteorological data to obtain the error determination result;

[0047] If the error determination result is lower than the preset accuracy requirement, then based on the performance index data, the resource allocation ratio is calculated to obtain a preliminary resource allocation scheme.

[0048] Based on the preliminary resource allocation scheme, resource scheduling optimization and load balancing are performed to obtain the computing resource allocation.

[0049] Preferably, the step of performing weather forecasting based on the computing resource allocation, the updated terrain influence weight matrix, and the spatial distribution feature set to obtain regional weather forecast results includes:

[0050] The computing resource configuration is executed, and task execution data is collected in real time;

[0051] Based on the task execution data, the updated terrain influence weight matrix, and the spatial distribution feature set, a weighted fusion is performed to obtain a fused feature vector set;

[0052] Based on the fused feature vector set, prediction calculations and real-time updates are performed to obtain regional weather forecast results.

[0053] Preferably, the step of generating a spatial distribution map based on the regional weather forecast results to obtain a spatial distribution map of local rainfall intensity, and performing differential corrections to obtain a regional weather forecast output, includes:

[0054] Obtain real-time weather data;

[0055] The rainfall probability distribution is extracted from the weather forecast results, and the rainfall intensity index of each grid is generated by grid division. The index that exceeds the intensity threshold is smoothed and adjusted. The spatial distribution map is generated by combining real-time meteorological data, thus obtaining the spatial distribution map of rainfall intensity.

[0056] Based on the spatial distribution map of rainfall intensity, the sea and land breeze influence factors are calculated, and combined with real-time meteorological data, a weighted fusion is performed to obtain the sea and land breeze modulation parameters.

[0057] Based on the sea and land breeze modulation parameters, spatial projection and optimized modulation are performed to output the final regional weather forecast results.

[0058] Secondly, the present invention provides a regional weather forecasting system based on a large meteorological model, comprising:

[0059] The weight matrix acquisition module is used to obtain the terrain influence weight matrix;

[0060] The spatial distribution feature module is used to input the terrain influence weight matrix into the pre-trained spatial distribution feature model, output an optimized spatial distribution feature set, and register it with spatiotemporal grid data to obtain the spatial distribution feature set of local weather.

[0061] The trend module is used to extract seasonal trends based on the spatial distribution feature set to obtain seasonal trends.

[0062] The correction coefficient module is used to perform seasonal deviation correction based on the seasonal variation trend to obtain the deviation correction coefficient;

[0063] The deep coupling module is used to update the terrain influence weight matrix based on the deviation correction coefficient to obtain the updated terrain influence weight matrix, and generate a parameter set to obtain the coupling parameter set for deep coupling of regional characteristics.

[0064] The resource configuration module is used to generate prediction results based on the set of coupling parameters. If the prediction results are lower than a preset accuracy threshold, then the computational resource configuration is generated to obtain the computational resource configuration.

[0065] The weather forecasting module is used to perform weather forecasting based on the computing resource configuration, the updated terrain influence weight matrix, and the spatial distribution feature set, and to obtain regional weather forecasting results.

[0066] The prediction output module is used to generate a spatial distribution map based on the regional weather prediction results, obtain a spatial distribution map of local rainfall intensity, and perform differential correction to obtain the regional weather prediction output.

[0067] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the regional weather forecasting method based on a large meteorological model as described above.

[0068] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the regional weather forecasting method based on a large meteorological model as described above.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] (1) In the process of processing the terrain influence weight matrix, by standardizing, spatially encoding and multi-layer convolution of the original terrain weight data, the gradient change and slope aspect information of the local terrain can be extracted step by step. The Kriging interpolation method is used to register with the time grid data to obtain the refined local weather spatial distribution characteristics, which helps to improve the basic prediction accuracy of complex terrain areas.

[0071] (2) In the seasonal change trend extraction stage, the main periodic components are extracted by performing frequency domain analysis on the spatial distribution feature set, and combined with time series coding and state memory processing, time series data that can reflect the characteristics of different seasons are obtained. With dimensionality reduction and interpolation completion, the responsiveness of the prediction results to seasonal changes can be significantly enhanced.

[0072] (3) In the computational resource allocation stage, error calculation is performed based on the preliminary forecast results and measured meteorological data. When the forecast accuracy is insufficient, the resource allocation ratio is calculated and the scheduling is optimized in combination with the performance index data to ensure that key computing tasks are given priority, thereby improving the forecast efficiency and stability in complex task scenarios.

[0073] (4) In the spatial distribution map generation and output stage, the rainfall probability in the prediction results is first divided into grids and the rainfall intensity index is calculated. Grids that exceed the threshold are smoothed. Then, the spatial distribution map of rainfall intensity is generated by combining real-time meteorological data. Based on this, the sea and land breeze influence parameters are calculated and projected onto the regional grid coordinate system to obtain a more realistic multidimensional weather prediction output. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the regional weather forecasting method based on a large meteorological model provided in the first embodiment of the present invention;

[0075] Figure 2 This is a schematic diagram of the structure of a regional weather forecasting system based on a large meteorological model provided in the second embodiment of the present invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Reference Figure 1 The first embodiment of the present invention provides a regional weather forecasting method based on a large meteorological model, comprising the following steps:

[0078] S11, obtain the terrain influence weight matrix;

[0079] S12, input the terrain influence weight matrix into the pre-trained spatial distribution feature model, output the optimized spatial distribution feature set, and register it with the spatiotemporal grid data to obtain the spatial distribution feature set of local weather;

[0080] S13, Based on the spatial distribution feature set, extract the seasonal variation trend to obtain the seasonal variation trend;

[0081] S14, Based on the aforementioned seasonal variation trend, perform seasonal deviation correction to obtain the deviation correction coefficient;

[0082] S15, Based on the deviation correction coefficient, the terrain influence weight matrix is ​​updated to obtain the updated terrain influence weight matrix, and a parameter set is generated to obtain the coupling parameter set of deep coupling of regional characteristics.

[0083] S16, Based on the set of coupling parameters, generate prediction results. If the prediction results are lower than a preset accuracy threshold, generate computing resource configuration to obtain computing resource configuration.

[0084] S17. Based on the computing resource allocation, the updated terrain influence weight matrix, and the spatial distribution feature set, perform weather forecasting to obtain regional weather forecast results.

[0085] S18. Based on the regional weather forecast results, a spatial distribution map is generated to obtain a spatial distribution map of local rainfall intensity, and differential corrections are performed to obtain the regional weather forecast output.

[0086] In step S11, the terrain influence weight matrix is ​​obtained;

[0087] It is worth noting that high-resolution digital elevation data (DEM) of the target area and meteorological observation data consistent with its time period are required. Elevation data can be downloaded from the SRTM public dataset through the National Geographic Information Public Service Platform, selecting the 30-meter resolution version to ensure sufficient clarity of terrain details. Meteorological observation data can be obtained from the National Meteorological Information Center or regional meteorological stations, including hourly records of rainfall, temperature, and wind speed, ensuring time coverage corresponds to the observation period of the elevation data. After obtaining the elevation data, the slope is calculated using the 3×3 neighborhood calculation method by reading the elevation value of each grid cell; that is, the elevation difference between the grid cell and its eight neighboring grid cells is calculated and converted into a slope percentage based on the horizontal distance. Slope aspect is calculated using the arctangent function. The azimuth angles are obtained, ranging from 0° to 360°; the roughness is calculated by representing the surface undulation as the standard deviation of elevation values ​​within the grid's neighborhood. Subsequently, these topographic feature data are interpolated to a unified 1-kilometer spatial resolution grid using the inverse distance weighted (IDW) method for later matching with meteorological data. Before interpolation, the meteorological observation data undergoes time synchronization processing, aligning hourly data from each meteorological station to the target time step and using Kriging interpolation to fill in spatially missing points, thus forming a spatiotemporally continuous meteorological element grid dataset. After completing spatiotemporal registration, two sets of data samples are prepared for each 1km × 1km grid: one set contains topographic feature values ​​(sine / cosine components of slope and aspect, roughness), and the other set contains meteorological element values ​​(such as precipitation, wind speed, air pressure, and humidity) from the same grid at the same time. The samples are organized in a two-dimensional "grid-time" manner: for example, selecting hourly data for 12 consecutive months will generate 8760 paired samples for a given grid. For each sample, x represents the topographic feature value of that grid, and y represents the meteorological observation value at that moment. Paired samples from all grids and all moments across the entire region are concatenated to form two sequences X and Y of equal length. Correlation calculation involves five steps: first, calculating the average of X and Y respectively; second, subtracting the corresponding average from each sample to obtain the mean-reduced sequence; third, calculating the standard deviation of the two mean-reduced sequences; fourth, multiplying the corresponding elements of the two mean-reduced sequences, summing the results, and averaging them to obtain the covariance; fifth, dividing the covariance by the standard deviation product to obtain the correlation coefficient, which ranges from -1 to 1. The closer the absolute value is to 1, the stronger the correlation. A threshold of 0.70 is set, meeting the meteorological statistics "strong correlation" standard, and is used to screen out significant influencing factors; those below the threshold are not included in subsequent calculations. When estimating the weights, meteorological elements are used as the output and the terrain features that "pass through" are used as the input. The parameters are first standardized, and then the coefficients are fitted using the least squares method. The absolute values ​​are scaled proportionally to sum to 1 to form the weight vectors of each grid. The results of the whole area are then spliced ​​together to generate the terrain influence weight matrix.

[0088] In step S12, the terrain influence weight matrix is ​​input into a pre-trained spatial distribution feature model, which outputs an optimized spatial distribution feature set. This set is then registered with spatiotemporal grid data to obtain a spatial distribution feature set of local weather, including:

[0089] Acquire time grid data;

[0090] The spatial distribution feature model is obtained by training a CNN neural network;

[0091] Through the input layer of the spatial distribution feature model, the terrain influence weight matrix is ​​normalized and spatially encoded to obtain a standard spatial weight tensor.

[0092] Through the convolutional layers of the spatial distribution feature model, multi-layer convolution operations are performed on the standard spatial weight tensor to extract local gradient and slope aspect features, thereby obtaining a preliminary spatial feature set.

[0093] The initial spatial feature set is subjected to max pooling dimensionality reduction through the pooling layer of the spatial distribution feature model to obtain the dimensionality-reduced spatial feature set.

[0094] The reduced spatial feature set is integrated through the fully connected layer of the spatial distribution feature model, and an optimized spatial distribution feature set is generated by nonlinear mapping through an activation function.

[0095] Based on the Kriging interpolation method and combined with the time grid data, missing values ​​are filled in the optimized spatial distribution feature set to obtain the spatial distribution feature set of local weather.

[0096] It is worth noting that the acquisition of time grid data involves: retrieving the time index of the target area for a recent period from the business database, with a time resolution of 1 hour, and the time range consistent with the subsequent prediction window; simultaneously reading the spatial grid description of the area, such as the row and column numbers and latitude and longitude grid of a 1 km × 1 km grid. The terrain influence weight matrix output in step S11 is spatially aligned to this grid; if the missing value ratio of individual grid points is less than 20%, it is filled using eight-neighbor inverse distance weighting; subsequently, standardization is performed for each channel by subtracting the channel mean and then dividing by the channel standard deviation, resulting in a multi-channel weight stack of size H × W × C. At this point, the input data has been aligned to the "time grid" at both the temporal and spatial scales.

[0097] The model training process is as follows: At least three years of historical samples are collected. The input is a multi-channel weight stack prepared as described above. The supervised target is a rainfall intensity distribution map corrected by radar and station network fusion at the same time, with the same resolution as the weight stack. The training set, validation set, and test set are split chronologically into three groups of 70%, 15%, and 15% respectively to avoid information leakage. The network structure adopts a stacked approach of "convolution—batch normalization—ReLU—max pooling," with a uniform 3×3 kernel and stride of 1. The first layer has 32 filters, the second layer has 64, and every two convolutional layers are followed by a 2×2 max pooling layer to halve the spatial size. At the end, the dimensionality-reduced features are flattened and mapped back to the same grid size as the target through a fully connected layer. The training objective is to minimize the sum of two terms: the mean squared error between the model output and the target rainfall map, and 1 minus the Pearson correlation coefficient between the two, with each coefficient weighted at 0.5. The optimizer used is Adam, and the learning rate is... The batch size is 16, and the maximum training time is 80 epochs. If the correlation coefficient on the validation set does not improve within 8 epochs, the training will stop early and roll back to the optimal weights. When the correlation coefficient on the test set reaches 0.78 or higher and the mean squared error is lower than 0.12, the parameters are fixed and used for online inference.

[0098] Spatial feature extraction is performed in the order of "convolution → nonlinearity → dimensionality reduction". The first layer uses a 3×3 sliding window to perform element-wise multiplication and addition of channel weights pixel by pixel, plus a bias, to obtain 32 response maps. Batch normalization is then applied to bring the values ​​back to a stable distribution, followed by ReLU to set negative responses to zero and preserve salient structures. The second layer repeats the same calculation but uses 64 filters to expand the receptive field and continuously capture terrain bands and windward and leeward transitions. Then, 2×2 max pooling is performed: the maximum value is taken for each 2×2 block, the size is halved, and the main ridge and valley lines are preserved. A combination of "convolution ×2 + pooling ×1" is stacked again to obtain an intermediate representation of H / 4×W / 4×128.

[0099] The dimensionality-reduced representation is flattened into vectors and fed into a fully connected mapping for feature integration. Before integration, a random deactivation factor of 0.2 is applied to the vectors to suppress overfitting; after integration, a ReLU is applied again to reduce weak contributions. Then, bilinear upsampling is used to restore the integrated result to the original raster size H×W. This results in an optimized spatial distribution feature map with each cell as a single value; one map is obtained for each time step, and a set is obtained for the entire time window, thus obtaining the optimized spatial distribution feature set.

[0100] Next, spatiotemporal registration and interpolation are performed on the "optimized spatial distribution feature set," with the object being the output of the previous step, not the original weights. Processing is done hourly by time grid: first, the feature set is aligned to the H×W business coordinate system by raster index; if some grid points are still missing values, ordinary Kriging interpolation is performed. The Kriging process consists of three steps: First, constructing a semi-variogram function. For each point to be estimated, known grid point pairs within a 5 km radius are collected, their distances and the squares of their numerical differences are calculated, and the average is calculated by binning. A spherical model is fitted to obtain the relationship between distance and semi-variogram. Second, solving for the weights. A Kriging equation system is constructed using the fitted semi-variogram function. The terms of the equation system are the pairwise semi-variograms between known points, and the right-hand side is the semi-variogram between the point to be estimated and each known point. The weights of each known point are solved, ensuring the sum of the weights is 1. Third, weighted evaluation. These weights are used to linearly combine the eigenvalues ​​of known grid points in the neighborhood to obtain the estimated value of the grid point to be estimated. The above process is repeated for the entire grid points every hour. To control for error, leave-one-out cross-validation is performed: each known point is temporarily treated as an unknown point for interpolation, and the mean square error is calculated; if the mean square error for a certain hour is higher than 0.10, the neighborhood radius is increased from 5 km to 8 km and the fit and solution are recalculated.

[0101] To ensure that these spatial representations accurately depict local weather distribution rather than noise, a final correlation gating is applied, still working on the "optimized spatial distribution feature set." Using the most recent 30 days as a sliding window, the hourly feature map is paired with the hourly measured rainfall map by grid point. The Pearson correlation coefficient is calculated using a five-step method (first, calculate the mean of the two maps; then, remove the mean; calculate the standard deviation of the two maps; calculate the covariance; finally, divide the covariance by the product of the standard deviations). Grid points with an absolute correlation coefficient of 0.75 or higher are marked as passing; grid points that do not pass are replaced by the average of the passing grid points in their eight neighborhoods. The threshold of 0.75 is chosen because meteorological operations consider a correlation above 0.70 as strong, and increasing it to 0.75 in mixed mountain and coastal areas further suppresses random correlations. After processing, a local weather spatial distribution feature set with a time step of 1 hour is obtained, perfectly consistent with the operational H×W grid.

[0102] In step S13, seasonal variation trends are extracted based on the spatial distribution feature set to obtain the seasonal variation trends, including:

[0103] Based on the spatial distribution feature set, Fourier transform is used to extract the main frequency component and calculate the main frequency energy ratio;

[0104] If the proportion of the dominant frequency energy is higher than the preset seasonal variability threshold, the spatial distribution feature set is input into the pre-trained time series model and a preliminary time series feature set is output.

[0105] The time series model is trained using an LSTM neural network;

[0106] Through the input layer of the time series model, the spatial distribution feature set is standardized and time series encoded to generate a standard time series tensor.

[0107] By using the hidden layer of the time series model, state memory features are extracted from the standard time series tensor to obtain the state memory feature sequence;

[0108] Through the output layer of the time series model, sequence mapping and feature optimization are performed on the state memory feature sequence to output a preliminary time series feature set;

[0109] The initial time series feature set is subjected to dimensionality reduction processing, and the main time series features are retained by principal component analysis to obtain the dimensionality-reduced time series feature set.

[0110] The reduced-dimensional time series feature set is subjected to nonlinear mapping to generate optimized time series features;

[0111] Based on the Kriging interpolation method and combined with the time grid data, missing values ​​are filled in for the optimized time series features to obtain the seasonal variation trend.

[0112] It is worth noting that the spatial distribution feature set is decomposed into different frequency components using Fast Fourier Transform, and the proportion of the squared amplitude to the total energy is calculated to obtain the dominant frequency energy proportion. If this proportion is higher than 60%, it is determined that the regional characteristics have stable seasonal fluctuations. This threshold is derived from the statistical analysis results of publicly available rainfall data from the National Meteorological Center from 2000 to 2020, which can effectively distinguish between stable annual cycles and short-term disturbances.

[0113] The specific training process for the time series model is as follows: Historical meteorological time series data for the target area over three to five consecutive years are acquired. This data includes variables such as temperature, relative humidity, rainfall, wind speed, and air pressure, with a temporal resolution consistent with the aforementioned spatial distribution feature set. For data from different meteorological stations or grid points, spatiotemporal registration is first performed, and missing values ​​are supplemented using Kriging interpolation to ensure the integrity of the time series. Subsequently, training samples are constructed using a sliding window approach, with a window length of 30 time steps and a step size of 1 time step. The first 29 time steps serve as the input sequence, and the 30th time step serves as the supervision label, thereby generating input-label corresponding sample pairs.

[0114] Secondly, all input variables are standardized according to the mean and standard deviation of the training set, and the standardized parameters are retained for the validation and testing phases to avoid information leakage. Multiple variables within the same grid are concatenated into a multidimensional input sequence in chronological order, and the training, validation, and test sets are divided according to a 70%, 15%, and 15% ratio to ensure that the datasets do not overlap in time.

[0115] The time series model consists of two layers of Long Short-Term Memory (LSTM) networks, each containing 64 units. A 20% random deactivation rate is introduced between layers to suppress overfitting. A fully connected mapping layer is connected to the output of the second layer to ensure its output dimension matches the supervision label. The training objective is to minimize the mean squared error between the predicted and true values. The Adam optimization algorithm is used, with an initial learning rate of 0.001. The learning rate is halved when the validation set error shows no decrease for five consecutive training epochs. The batch size is set to 64, the maximum training epoch is 200 epochs, and training is stopped early when the validation set error shows no improvement for ten consecutive training epochs, retaining only the network parameters with the lowest validation set error as the final weights.

[0116] After the model training is completed, the input data of the spatial distribution feature set is sent to the input layer. The role of the input layer is to receive the temporally distributed spatial distribution features and perform data standardization processing, that is, to subtract the mean of the feature in the entire time series from the data at each time step, and then divide by the standard deviation, so as to eliminate the differences in units and scales and generate a standardized time series tensor with consistent numerical distribution that is easy to calculate later.

[0117] After entering the hidden layer, the input sequence is processed step by step using Long Short-Term Memory (LSTM) units. At each time step, the hidden layer sequentially performs a forget gate operation to determine the proportion of historical information to retain, performs an input gate operation to update the feature state of the current time step, calculates new memory unit values ​​using candidate states, and finally combines the output gate to pass the current unit state to the next time step. This process can simultaneously preserve long-term dependency information and short-term change patterns, thus obtaining a state memory feature sequence containing temporal dependencies.

[0118] In the output layer, a fully connected mapping is used to project the state memory feature sequence onto the target feature dimension, thereby realizing the recombination and optimization of features. This allows the initial time series feature set output to retain the original time-series structure while enhancing its relevance to the target analysis task.

[0119] The initial time series feature set is dimensionality reduced by using principal component analysis (PCA) to preserve key time-series features. Specifically, the initial feature set is first zero-mean normalized column-wise to eliminate numerical biases between features. Then, the covariance matrix of each feature is calculated to reflect their linear correlation. Next, eigenvalue decomposition is performed on the covariance matrix to obtain a set of eigenvectors and their corresponding eigenvalues. The eigenvalues ​​are arranged in descending order, and the cumulative variance contribution rate is calculated. The top few eigenvectors with cumulative contribution rates reaching a preset proportion (e.g., 90%) are selected, and the original data is projected onto these eigenvectors to form a new low-dimensional dataset—the dimensionality-reduced time series feature set. This approach preserves key time-series variation information while reducing redundant dimensions and subsequent computational load.

[0120] Subsequently, a nonlinear mapping is applied to the dimensionality-reduced time series feature set to enhance its feature representation capability. During execution, the dimensionality-reduced features are input into a fully connected structure, and a new feature vector is formed through a linear combination of weights and inputs. Nonlinear functions such as ReLU are then applied to transform the output element-wise, thereby introducing a nonlinear feature change pattern. This allows the optimized time series features output to better characterize the seasonal fluctuations in the original data.

[0121] Finally, based on the Kriging interpolation method combined with time grid data, missing values ​​are filled in for the optimized time series features. Specifically, a semi-variogram is calculated based on the feature values ​​of existing grid points to fit a spatial autocorrelation relationship; this relationship is then used to assign weights to neighboring observation points for missing points, and the interpolated value is obtained by weighted summation. This method can minimize prediction errors while maintaining spatial continuity, generating a complete dataset of seasonal variation trends.

[0122] In step S14, based on the seasonal variation trend, seasonal deviation correction is performed to obtain the deviation correction coefficient, including:

[0123] Acquire real-time environmental data and seasonal feature databases;

[0124] Based on the mean filtering method, the real-time environmental data is smoothed to obtain a smoothed environmental data set;

[0125] The difference between the environmental data set and the seasonal feature database is calculated to obtain the deviation value. The deviation is then corrected by combining the seasonal change trend to obtain an initial deviation value set.

[0126] If the absolute value of any parameter in the initial deviation value set exceeds the corresponding preset deviation threshold, the offset parameter is adjusted to obtain the adjusted offset parameter set, and an optimized deviation correction coefficient is generated through a linear transformation method.

[0127] It is worth noting that, firstly, environmental data for the current moment is collected from temperature, humidity, and barometric pressure sensors. A seasonal feature database is then established. Specifically, historical observation data from the same type of sensor in the target area over the past 5–10 years are selected and grouped by month and meteorological season (Spring (March–May), Summer (June–August), Autumn (September–November), and Winter (December–February). The mean and standard deviation of temperature, relative humidity, and barometric pressure for each season are calculated as the baseline statistics for that season. For missing data segments, linear interpolation is first performed, and then outliers are removed using a box-ring method to ensure data consistency within the database. The box-ring method is specifically set to a 1.5x IQR rule, meaning that observations below the first quartile minus 1.5x the interquartile range, or above the third quartile plus 1.5x the interquartile range, are identified as outliers and removed.

[0128] To reduce short-term fluctuations, the collected temperature, humidity, and air pressure data were subjected to mean filtering within a fixed 5-minute window. This involved averaging the current value and the values ​​from the previous 4 minutes to obtain a smoothed environmental data set. Subsequently, the smoothed temperature, humidity, and air pressure values ​​were subtracted from the baseline mean in the seasonal feature database to obtain the three component deviations. These deviations were then standardized according to the baseline standard deviation to ensure comparability under a unified dimension, thus forming an initial set of deviation values.

[0129] If the absolute value of any parameter in the initial deviation value set exceeds a preset deviation threshold, the offset parameter adjustment process begins. The deviation thresholds are determined through long-term observation and statistical analysis, with the temperature threshold set at ±2°C, the humidity threshold at ±5%, and the air pressure threshold at ±3 hPa. These values ​​are based on the following: when the temperature deviation exceeds 2°C, the difference between the local observation point and the seasonal baseline reaches a significant level affecting short-term meteorological trend judgment; when the humidity deviation exceeds 5%, it will cause significant errors in precipitation forecasting and evaporation calculation; and when the air pressure deviation exceeds 3 hPa, it corresponds to a substantial movement of the weather system.

[0130] In parameter adjustment, the lateral offset angle is determined by constructing a set of surrounding measuring points centered on the current measuring point to be corrected. This set consists of valid observation points located within a 10-kilometer radius of the central measuring point and at similar altitudes (elevation difference not exceeding 100 meters). These observation points include data from national meteorological stations and sensor nodes deployed in the field. This definition ensures the representative spatial consistency of the measuring points involved in the correction. Subsequently, the spatial differences between the central measuring point and each surrounding measuring point are calculated. The geographic coordinate differences are converted from latitude and longitude points in the WGS-84 coordinate system to planar distances in meters using the Haversine spherical distance formula. The reciprocal of the distance is used as a weight to weight the observations of the surrounding measuring points, ultimately forming the lateral offset angle to correct for the spatial influence on planar position.

[0131] In the calculation of the longitudinal offset angle, the difference in altitude between the central measuring point and the surrounding measuring points is extracted by combining the meteorological observation records of the same season within 5 to 10 years. Based on empirical climate gradient parameters (such as the temperature lapse rate set at 0.6°C / 100m), the changes in meteorological variables caused by elevation are quantitatively corrected to compensate for the prediction deviation caused by the vertical topographic differences.

[0132] The time difference weighting is achieved by performing cross-correlation analysis on the smoothed time series and the baseline time series to find the time delay corresponding to the peak correlation coefficient, and then normalizing it to a weighting factor between 0 and 1. The results of these three adjustments are combined to form a set of offset parameters that includes the horizontal offset angle, the vertical offset angle, and the time difference weighting.

[0133] For the contour difference values ​​corresponding to the offset parameter set, the Kriging interpolation method is used. The weighting coefficients of each prediction point are calculated by using the spatial autocorrelation and semivariogram fitting results. Spatial registration is performed on the missing or abnormal grid point data to obtain the registered contour difference values. Then, the deviation correction coefficient is calculated. The initial deviation value set consists of the standardized temperature, humidity, and air pressure difference values ​​of the current measuring point. For example, the temperature deviation of measuring point A is +1.5, the humidity deviation is −0.8, and the air pressure deviation is +2.1, all of which have been normalized according to their respective seasonal standard deviations. The offset parameter set consists of the lateral offset angle, the longitudinal offset angle, and the time difference weight. The lateral offset angle is calculated from the horizontal distance between measuring point A and its three surrounding measuring points B, C, and D. For example, the distances of B, C, and D from A are 2km, 4km, and 8km, respectively, so their reciprocals are 0.5, 0.25, and 0.125, respectively. After normalization, the weights are 0.57, 0.29, and 0.14, which are used to weight the correction intensity of the current point in the spatial lateral distribution. The longitudinal offset angle is calculated based on the measured... The average elevation difference between point A and surrounding measuring points is calculated. For example, if A is 500 meters and the surrounding points are 300 meters apart, the elevation difference is 200 meters. If the maximum impact elevation difference is set to 400 meters, the vertical weight is 200 divided by 400, resulting in 0.5, indicating that the point has a moderate degree of vertical correction necessity. The time difference weight is calculated by performing cross-correlation analysis on the smoothed time series and the baseline time series. If the maximum correlation occurs at a lag of 6 minutes, and considering a maximum allowable lag of 10 minutes, the normalized time difference weight is 0.6. The contour difference value reflects the degree of undulation of the measuring point in the terrain surface. It is calculated using the semi-variogram obtained by Kriging interpolation. This value is in the median of the target area, for example, 1.0. If the maximum difference value in the area is 2.0, the contour correction coefficient is 0.5. During the fusion process, the system weights each of the four correction factors with each component of the initial deviation value. For example, if the temperature deviation is +1.5, the system multiplies it by 0.57, 0.5, 0.6, and 0.5 respectively, resulting in four values ​​of +0.855, +0.75, +0.9, and +0.75. The average of these four values ​​is then taken, resulting in an optimized temperature correction value of +0.814. The humidity and air pressure deviation components are processed in the same way, each combined with its corresponding four weighting factors. It is important to note that all weighting coefficients are derived from the physical distance, elevation difference, time lag, and terrain contour indicators of the actual measurement points. The weights are set between 0 and 1 and maintain overall weight balance to avoid error shifts caused by bias towards a single factor. Through these steps, the system integrates the spatial, vertical, and temporal information of various influencing factors and generates a set of optimized deviation correction coefficients for subsequent prediction and modulation after targeted weighting correction of the original deviation.

[0134] In step S15, based on the deviation correction coefficient, the terrain influence weight matrix is ​​updated to obtain an updated terrain influence weight matrix, and a parameter set is generated to obtain a coupling parameter set for deep coupling of regional characteristics, including:

[0135] Acquire regional prior data and regional characteristic data;

[0136] Based on the deviation correction coefficient, the terrain influence weight matrix is ​​weighted and adjusted to obtain the updated initial weight matrix;

[0137] Based on the initial weight matrix and the prior data of the region, local sensitivity factors are generated to obtain a set of local sensitivity factors.

[0138] If the local sensitive factor set has preset sensitivity judgment conditions, the initial weight matrix is ​​gradually adjusted to obtain the updated terrain influence weight matrix, and combined with the regional characteristic data, deep coupling calculation is performed to obtain the coupling parameter set of deep coupling of regional characteristics.

[0139] It is worth noting that prior geological data and regional characteristic data of the target area were obtained. Prior geological data, including soil type, rock strata distribution, groundwater level, and slope aspect and gradient, were obtained from regional geological survey results and historical mapping data; regional characteristic data, including surface roughness, location of the land-sea boundary, vegetation coverage, and local climate parameters, were obtained from remote sensing image analysis, long-term observations from meteorological stations, and field measurements.

[0140] Subsequently, the deviation correction coefficient is applied to the terrain influence weight matrix, and the elements of the matrix are adjusted by weight to obtain the updated initial weight matrix. During the weight calculation, the adjustment range of the weights is determined based on the absolute value and direction of the deviation correction coefficient. For example, when the temperature and humidity deviation of a certain grid point is positive and the amplitude is large, the weight of the corresponding climate factor is increased, while the weight of the construction condition factor is appropriately reduced, thereby reflecting the difference in the importance of environmental changes to the prediction parameters.

[0141] After obtaining the initial weight matrix, a set of local sensitivity factors is calculated by combining regional characteristic data. Specifically, for each weight element, a normalization relationship is established with the corresponding factor in the regional characteristic data (such as slope, roughness, and land-sea boundary distance) to calculate the local sensitivity score. The normalization process uses an interval scaling method to map the original value of each factor to a standard range between 0 and 1. Taking the slope factor as an example, if the slope value in the region ranges from 3° to 30°, the original slope value is transformed into a normalized value by linear transformation, using the method of "subtracting the minimum value from the current value and then dividing by the range". For example, if the slope of an observation point is 21°, the normalized value is (21-3) / (30-3) = 0.67. After completing the normalization process, the weighted calculation is performed according to the set sensitivity weights to obtain the comprehensive sensitivity score. Among them, the weight of the slope factor is 0.35, the roughness factor is 0.25, the land-sea boundary distance factor is 0.15, and the weight of the remaining factors is 0.25. The weighting process involves multiplying the normalized value of each factor by its corresponding weight and summing all the results. For example, if a point has a normalized slope of 0.67, a roughness of 0.48, and a land-sea boundary distance of 0.30, and the total normalized value of the remaining factors is 0.55, then the comprehensive sensitivity score is calculated as follows: 0.67×0.35 +0.48×0.25 + 0.30×0.15 + 0.55×0.25 = 0.233 + 0.12 + 0.045 + 0.1375 = 0.5355.

[0142] A sensitivity score threshold is set to determine whether a point should enter the weight fine-tuning process. If the overall score is greater than or equal to 0.75, or any single score is greater than or equal to 0.80, the point is determined to be a high-sensitivity feature region and enters the subsequent iterative fine-tuning process; otherwise, it is skipped, and the current weight structure remains unchanged. The above threshold value can effectively identify high-influence areas while avoiding excessive weight adjustment in low-influence areas.

[0143] Then, using the current initial weight matrix as a baseline, the weights corresponding to various geological, climatic, and construction condition factors are fine-tuned according to the proportion of locally sensitive factors. For example, if the slope factor score is higher than the threshold, the proportion of terrain-related weights is increased and the weight of irrelevant factors is reduced during the adjustment. Terrain-related weights refer to factors directly related to the regional landform, including slope, surface relief, and terrain elevation; while irrelevant factors refer to climatic or environmental parameters with weak correlation to terrain response, such as average wind speed and solar radiation. The fine-tuning method adopts an iterative optimization approach, with each round of adjustment controlled between 0.5% and 1% of the original weight value. The maximum number of iterations is set to 20, or the process automatically stops when the adjustment range is less than 0.1% for two consecutive rounds. Taking a grid point in a mountainous area as an example, if the slope sensitivity score is 0.82, which exceeds the threshold of 0.80, the weight of this factor can be increased from the original 0.12 to 0.13; at the same time, if the wind speed sensitivity score is only 0.22, its weight can be reduced from the original value of 0.08 to 0.075 in order to maintain the balance of the total weight of the matrix.

[0144] After fine-tuning, the updated terrain influence weight matrix is ​​compared item by item with the regional characteristic data. Correlation calculation methods are used to determine the coupling strength between different factor pairs. Specifically, a correlation judgment method based on the changing trends between spatial data columns is adopted: the changing trend of a parameter (such as slope) at different geographic grid points is compared with the changing trend of another parameter (such as rainfall) at the same grid points. First, the mean values ​​of both sets of parameters are calculated across all grid points. Then, for each grid point, the difference between the parameter and its mean value (i.e., the center offset) is calculated. Subsequently, the offsets of all grid points are multiplied pairwise and summed to obtain the total value of the coordinated change. Finally, this is compared with the product of the sum of the squares of the offsets of the two sets of parameters to determine whether the direction and intensity of their fluctuations are consistent. If the changing trends of the two are highly consistent—that is, when the slope value increases, the rainfall also increases—and this consistency holds true in most grid points, it indicates that the two are tightly coupled.

[0145] For example, in a mountainous area, calculations of slope and rainfall data from 200 grid points revealed that slope and rainfall showed a synchronous upward trend in most grid points, with a highly consistent coupling trend. The correlation coefficient between the two was found to be approximately 0.88. Based on this value, this parameter pair was identified as a highly correlated parameter. In the final deep coupling weight calculation, its combined weight can be increased by more than 10%, thus more fully reflecting the significant impact of the coordinated changes in topography and climate on the prediction results.

[0146] In step S16, a prediction result is generated based on the coupling parameter set. If the prediction result is lower than a preset accuracy threshold, a computing resource configuration is generated to obtain the computing resource configuration, including:

[0147] Acquire performance index data and measured meteorological data for the target area;

[0148] Based on the set of coupling parameters and combined with the regional characteristic data, weather forecasting is performed to obtain preliminary forecast meteorological data for the target area.

[0149] The prediction error is calculated by comparing the preliminary meteorological data with the measured meteorological data to obtain the error determination result;

[0150] If the error determination result is lower than the preset accuracy requirement, then based on the performance index data, the resource allocation ratio is calculated to obtain a preliminary resource allocation scheme.

[0151] Based on the preliminary resource allocation scheme, resource scheduling optimization and load balancing are performed to obtain the computing resource allocation.

[0152] It is worth noting that performance index data and measured meteorological data of the target area are retrieved from the database. The performance index data includes historical forecast accuracy records, computation time, node CPU utilization, and memory usage. Measured meteorological data can be obtained from the National Meteorological Information Center, including hourly rainfall, wind speed, air pressure, temperature, and other variables, and is uniformly organized according to a 1-hour time step. To ensure data consistency, all meteorological data must be spatially aligned to a 1-kilometer resolution grid and temporally completely overlap with the forecast window generated in the previous step. After loading, all data undergoes missing data filling processing, using Kriging interpolation to fill in grid point gaps. The spatial neighborhood radius during interpolation is set to 5 kilometers, and a spherical model is used for fitting the variogram. If the cross-validation error exceeds 0.1, the radius is increased to 8 kilometers and recalculated.

[0153] During the prediction process, the system uses an hourly time step and each square kilometer as the calculation unit. For each grid point, it extracts five standardized features: terrain slope, surface roughness, soil moisture, altitude, and vegetation cover. These features are then weighted and combined according to pre-determined fixed weights to calculate predicted values ​​for rainfall, temperature, wind speed, and other factors. For example, in rainfall prediction, the system primarily considers soil moisture and slope; in wind speed prediction, it mainly relies on surface roughness and altitude. Taking grid A123 as an example, its soil moisture is 0.8, slope is 0.6, roughness is 0.3, altitude is 0.5, and vegetation cover is 0.4. The system weights and combines these features according to the rainfall weight ratio, outputting a predicted rainfall value of 5.2 mm for that moment. The predicted values ​​from all grids are then compiled into a unified meteorological forecast.

[0154] Error calculations were performed between predicted and observed meteorological data. The absolute and relative errors of each variable were calculated using a grid-by-grid interpolation method. The error calculation method involved subtracting the observed value from the predicted value for each grid to obtain an error value sequence, from which the mean squared error (MSE) and mean absolute error (MAE) were calculated. The MSE is used to measure prediction stability, while the MAE is used to assess the overall bias level. Both are accumulated hourly and compared to a preset accuracy threshold.

[0155] The accuracy thresholds were set with reference to the error statistics and evaluation standards of similar regions over the past five years. Taking rainfall as an example, a prediction error of less than ±1.5 mm / h was considered to meet the requirements for short-term heavy rainfall forecasting; the temperature error threshold was set at ±1°C, and the wind speed threshold was set at ±0.5 m / s. If the errors of all major variables did not exceed the above thresholds, the prediction accuracy was deemed acceptable; otherwise, the resource optimization process was initiated.

[0156] If the error is determined to be insufficient to meet accuracy requirements, the resource allocation ratio will be readjusted based on the measured performance data. The specific adjustment process includes analyzing performance data such as task queue length, CPU utilization, and average task response time to determine the resource stress of each node. In practice, the system will prioritize nodes under high load. For example, if a node's CPU utilization exceeds 85% and the average response time of its corresponding task exceeds 100 milliseconds, the system will consider this node to have a performance bottleneck and increase its resource allocation ratio by approximately 10% to enhance its processing capacity. Simultaneously, the resource allocation ratio will be reduced by a moderate proportion from nodes with lower loads to ensure that overall resources are not wasted.

[0157] The third step is to generate a preliminary allocation plan based on the adjusted resource allocation ratio. This plan is saved in the format of "Compute Node ID → Allocated Resource Ratio", which has a clear structure and is easy for scheduling tools to parse.

[0158] Subsequently, the initial allocation scheme is combined with regional characteristic data to optimize resource scheduling. If a region has complex terrain and highly variable weather, such as a mountainous area where the coast meets the mountains, high-performance computing nodes and larger memory resources are prioritized to support the real-time computational needs of complex terrain factors. After the scheduling scheme is generated, a load balancing strategy is used to allocate the final computing tasks, ensuring that the CPU utilization of all computing nodes does not exceed 90% and the response time does not exceed 150ms, thereby generating the final computing resource configuration.

[0159] In step S17, weather forecasting is performed based on the computing resource configuration, the updated terrain influence weight matrix, and the spatial distribution feature set to obtain regional weather forecast results, including:

[0160] The computing resource configuration is executed, and task execution data is collected in real time;

[0161] Based on the task execution data, the updated terrain influence weight matrix, and the spatial distribution feature set, a weighted fusion is performed to obtain a fused feature vector set;

[0162] Based on the fused feature vector set, prediction calculations and real-time updates are performed to obtain regional weather forecast results.

[0163] It is worth noting that various computing resources are scheduled to complete parallel task allocation, and task instructions are issued to edge computing nodes and core servers, including preprocessing of input data, generation of fused features, and prediction output. Subsequently, based on the task type and data update frequency defined in the current scheduling plan, the latest terrain influence weight matrix and the spatial distribution feature set at the corresponding time are extracted, such as the slope, roughness, soil moisture, and wind speed distribution features up to the current hour.

[0164] Channel alignment and scale normalization are performed on the two datasets to ensure consistent input dimensions. Then, a point-by-point fusion operation is performed using weighted aggregation and feature combination. For example, at a certain grid point, the terrain weight is 0.6, humidity is 0.75, and wind speed is 3 meters per second. By setting a weighting coefficient, the comprehensive feature of this point is calculated as 0.6 multiplied by 0.75 plus a normalized value of 0.4 multiplied by 3, resulting in a fused feature vector. Consistent operations are performed on all grid points to form a complete fused feature vector set.

[0165] After generating the fused feature vector set, the system extracts preliminary predicted values ​​such as temperature, rainfall intensity, and rainfall probability for each grid cell using a pre-established meteorological variable fitting relationship table. This relationship table is derived from a large amount of measured data over the past three years, divided into 100 equal-width ranges based on the fused feature value from 0.00 to 1.00. Each range records the statistical mean of the corresponding measured variable in the past. Taking a grid cell with a fused feature value of 0.68 in the current period as an example, the system searches for range number 68 and reads the corresponding historical values: temperature 16.5 degrees Celsius, rainfall intensity 1.6 mm / hour, and rainfall probability 72%. These three data points constitute the preliminary prediction results for that grid cell and possess statistical representativeness consistent with historical meteorological models.

[0166] To further improve the accuracy of the forecast response, the system reads the measured meteorological data of the grid in the most recent hour and combines it with the average error of similar forecasts and measurements over the past seven days to construct a correction factor. The fitting factor is calculated by statistically analyzing the differences between the historical forecast values ​​and measured values ​​of a certain grid point over the past seven days, and taking the average as the factor correction amount. For example, if the fusion feature value of a certain grid is 0.68, the corresponding historical average temperature deviation is +0.9 degrees Celsius and the rainfall intensity deviation is +0.2 mm / hour. If the initial forecast value for this period is 16.5 degrees Celsius and 1.6 mm / hour, the corrected value is 17.4 degrees Celsius and 1.8 mm / hour. The probability of rainfall remains 72%. The corrected meteorological variables are organized and summarized according to the grid number to form a consistent forecast data set, and the regional weather forecast result for the current moment is output.

[0167] In step S18, based on the regional weather forecast results, a spatial distribution map is generated to obtain a spatial distribution map of local rainfall intensity, and differential corrections are performed to obtain the regional weather forecast output, including:

[0168] Obtain real-time weather data;

[0169] The rainfall probability distribution is extracted from the weather forecast results, and the rainfall intensity index of each grid is generated by grid division. The index that exceeds the intensity threshold is smoothed and adjusted. The spatial distribution map is generated by combining real-time meteorological data, thus obtaining the spatial distribution map of rainfall intensity.

[0170] Based on the spatial distribution map of rainfall intensity, the sea and land breeze influence factors are calculated, and combined with real-time meteorological data, a weighted fusion is performed to obtain the sea and land breeze modulation parameters.

[0171] Based on the sea and land breeze modulation parameters, spatial projection and optimized modulation are performed to output the final regional weather forecast results.

[0172] It's worth noting that the system extracts the rainfall probability field from each grid in the regional weather forecast results and uses a fixed-interval latitude and longitude grid division method to divide the entire target area into several regular grid units. Each grid corresponds to an initial rainfall probability value, expressed as a percentage. Next, the system multiplies this rainfall probability by a fixed mapping coefficient to convert it into a rainfall intensity index. For example, if the rainfall probability of a certain grid is 70%, multiplying it by a coefficient of 0.05 results in an initial rainfall intensity value of 3.5 mm / hour. To avoid spatial discontinuities caused by abrupt changes in rainfall intensity, the system iterates through each grid, determining whether the intensity difference between it and its eight neighboring grids exceeds a set threshold of 2 mm / hour. If it does, a sliding smoothing process is performed on that point, averaging the current point with the average of its four neighboring grids. Through this processing, a spatially smoothed rainfall intensity distribution map is obtained.

[0173] After the rainfall intensity distribution map is constructed, the system further calculates the sea-land breeze influence factor based on the geographical attributes of the grid locations. The system first reads the base map data and topographic annotation map to determine whether each grid is a coastal area. The specific calculation method is as follows: grids less than 15 kilometers from the coastline are assigned an initial sea-land breeze influence factor of 1.0; grids between 15 and 30 kilometers have a factor of 0.8; and grids more than 30 kilometers have a factor of 0.5. Subsequently, the factor is weighted and corrected by incorporating wind direction and wind speed fields from measured meteorological data. For example, if a grid is located 8 kilometers from the coast with an initial factor of 1.0, a current wind speed of 12 meters per second, and a wind direction pointing inland, a correction of +0.2 is applied according to the set rules, resulting in an influence factor of 1.2. If the wind direction is offshore, the factor is reduced to 0.85 using a reverse weighting coefficient.

[0174] After completing the above factor corrections, the system performs point-weighted processing on the rainfall intensity spatial map based on the final sea-land breeze modulation factor for each grid. The final rainfall intensity value for each grid is equal to its smoothed rainfall intensity multiplied by the corresponding sea-land breeze factor. For example, if a grid has a rainfall intensity of 3.5 mm / h and a corresponding modulation factor of 1.2, then the final rainfall intensity at that point is 4.2 mm / h. The system repeats the above operation for all grid areas to generate an updated rainfall intensity spatial map.

[0175] Finally, the system maps the map back to the regional weather grid coordinate system and outputs the final regional weather forecast result containing complete fields such as temperature, rainfall intensity, and rainfall probability for each grid.

[0176] Reference Figure 2 The second embodiment of the present invention provides a regional weather forecasting system based on a large meteorological model, comprising:

[0177] The weight matrix acquisition module is used to obtain the terrain influence weight matrix;

[0178] The spatial distribution feature module is used to input the terrain influence weight matrix into the pre-trained spatial distribution feature model, output an optimized spatial distribution feature set, and register it with spatiotemporal grid data to obtain the spatial distribution feature set of local weather.

[0179] The trend module is used to extract seasonal trends based on the spatial distribution feature set to obtain seasonal trends.

[0180] The correction coefficient module is used to perform seasonal deviation correction based on the seasonal variation trend to obtain the deviation correction coefficient;

[0181] The deep coupling module is used to update the terrain influence weight matrix based on the deviation correction coefficient to obtain the updated terrain influence weight matrix, and generate a parameter set to obtain the coupling parameter set for deep coupling of regional characteristics.

[0182] The resource configuration module is used to generate prediction results based on the set of coupling parameters. If the prediction results are lower than a preset accuracy threshold, then the computational resource configuration is generated to obtain the computational resource configuration.

[0183] The weather forecasting module is used to perform weather forecasting based on the computing resource configuration, the updated terrain influence weight matrix, and the spatial distribution feature set, and to obtain regional weather forecasting results.

[0184] The prediction output module is used to generate a spatial distribution map based on the regional weather prediction results, obtain a spatial distribution map of local rainfall intensity, and perform differential correction to obtain the regional weather prediction output.

[0185] It should be noted that the regional weather forecasting system based on a large meteorological model provided in this embodiment of the invention is used to execute all the process steps of the regional weather forecasting method based on a large meteorological model in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0186] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a regional weather forecasting program based on a large meteorological model. When the processor executes the computer program, it implements the steps described in the various embodiments of the regional weather forecasting method based on a large meteorological model, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the trend change module.

[0187] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0188] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0189] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0190] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0191] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0192] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0193] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for regional weather prediction based on a meteorological large model, characterized in that, The method comprises the following steps: obtaining a terrain influence weight matrix; inputting the terrain influence weight matrix into a pre-trained spatial distribution feature model to output an optimized spatial distribution feature set, and performing space-time registration to obtain a spatial distribution feature set of local weather; extracting a seasonal change trend according to the spatial distribution feature set to obtain the seasonal change trend; based on the seasonal change trend, performing seasonal bias correction to obtain a bias correction coefficient, including: obtaining real-time environmental data and a seasonal feature library; performing smoothing processing on the real-time environmental data to obtain a smoothed environmental data set; calculating the bias value of the environmental data set and the seasonal feature library, and combining the seasonal change trend to perform bias correction to obtain an initial bias value set; if the absolute value of any parameter in the initial bias value set exceeds the corresponding preset bias threshold, adjusting the offset parameter to obtain an adjusted offset parameter set, and generating an optimized bias correction coefficient through linear transformation; wherein the offset parameter set is composed of a horizontal offset angle, a vertical offset angle and a time difference weight; a center point is constructed around a current to-be-corrected measuring point, the spatial difference between the center point and each surrounding measuring point is calculated, the geographical coordinate difference is converted into a plane distance in meters by using the Haversine spherical distance formula, and the reciprocal of the distance is used as the weight to add up the observation values of the surrounding measuring points to form the horizontal offset angle; in the calculation of the vertical offset angle, the altitude difference between the center point and the surrounding measuring points is extracted, and the change of the meteorological variable caused by the elevation is quantitatively corrected according to the empirical climate gradient parameter to compensate for the prediction bias caused by the vertical terrain difference; the time difference weight is obtained by cross-correlation analysis of the smoothed time series and the reference time series to find the time delay corresponding to the correlation coefficient peak, and normalizing it to a weight factor between 0 and 1; based on the bias correction coefficient, updating the weight of the terrain influence weight matrix to obtain an updated terrain influence weight matrix, and performing coupling calculation on the updated terrain influence weight matrix to obtain a coupling parameter set of regional characteristic depth coupling; according to the coupling parameter set, generating a prediction result, if the prediction result is lower than a preset precision threshold, generating a computing resource configuration to obtain a computing resource configuration; based on the computing resource configuration, the updated terrain influence weight matrix and the spatial distribution feature set, performing weather prediction to obtain a regional weather prediction result; based on the regional weather prediction result, generating a spatial distribution map to obtain a spatial distribution map of local rainfall intensity, and performing differential correction to obtain a regional weather prediction output. 2.The regional weather prediction method based on a meteorological model according to claim 1, wherein, The method comprises the following steps: obtaining time grid data; the spatial distribution feature model is trained by a CNN neural network; The terrain influence weight matrix is subjected to data normalization and spatial coding through the input layer of the spatial distribution feature model, to obtain a standard spatial weight tensor; The standard spatial weight tensor is subjected to multi-layer convolution operation through the convolution layer of the spatial distribution feature model, to extract local gradient and slope direction features, to obtain a preliminary spatial feature set; The preliminary spatial feature set is subjected to maximum pooling dimension reduction processing through the pooling layer of the spatial distribution feature model, to obtain a dimension-reduced spatial feature set; The dimension-reduced spatial feature set is integrated through the full connection layer of the spatial distribution feature model, and is subjected to non-linear mapping through an activation function, to generate an optimized spatial distribution feature set; Based on the Kriging interpolation method, the optimized spatial distribution feature set is subjected to missing value filling in combination with the time grid data, to obtain a spatial distribution feature set of local weather. 3.The regional weather prediction method based on a meteorological model according to claim 2, wherein, The seasonal change trend is extracted from the spatial distribution feature set, to obtain a seasonal change trend, including: The main frequency component is extracted by Fourier transform from the spatial distribution feature set, to calculate a main frequency energy proportion; If the main frequency energy proportion is higher than a preset seasonal change threshold, the spatial distribution feature set is input into a pre-trained time series model, to output a preliminary time series feature set; The time series model is trained by an LSTM neural network; The spatial distribution feature set is subjected to data standardization and time series coding through the input layer of the time series model, to generate a standard time series tensor; The standard time series tensor is subjected to state memory feature extraction through the hidden layer of the time series model, to obtain a state memory feature sequence; The state memory feature sequence is subjected to sequence mapping and feature optimization through the output layer of the time series model, to output a preliminary time series feature set; The preliminary time series feature set is subjected to dimension reduction processing, and main component analysis method is used to retain main time sequence features, to obtain a dimension-reduced time series feature set; The dimension-reduced time series feature set is subjected to non-linear mapping, to generate an optimized time series feature; Based on the Kriging interpolation method, the optimized time series feature is subjected to missing value filling in combination with the time grid data, to obtain a seasonal change trend. 4.The method of claim 1, wherein, The terrain influence weight matrix is updated based on the bias correction coefficient to obtain an updated terrain influence weight matrix, and the updated terrain influence weight matrix is subjected to coupling calculation, to obtain a coupling parameter set of regional characteristic deep coupling, including: Regional prior data and regional characteristic data are acquired; The terrain influence weight matrix is subjected to weighted adjustment based on the bias correction coefficient, to obtain an updated initial weight matrix; Local sensitive factors are generated based on the updated initial weight matrix and the regional prior data, to obtain a local sensitive factor set; If the local sensitivity factor set meets the preset sensitivity judgment condition, the updated initial weight matrix is gradually adjusted to obtain an updated terrain influence weight matrix, and deep coupling calculation is performed in combination with the regional characteristic data to obtain a coupling parameter set of regional characteristic deep coupling. 5.The regional weather prediction method based on a meteorological model according to claim 4, wherein, The prediction result generation is performed according to the coupling parameter set, and if the prediction result is lower than a preset precision threshold, the calculation resource configuration generation is performed to obtain a calculation resource configuration, including: Obtain performance index data and target area measured meteorological data; Based on the coupling parameter set, in combination with the regional characteristic data, weather prediction is performed to obtain preliminary predicted meteorological data of the target area; The preliminary predicted meteorological data and the measured meteorological data are used for prediction error calculation to obtain an error determination result; If the error determination result is lower than a preset accuracy requirement, resource allocation ratio calculation is performed based on the performance index data to obtain a preliminary resource configuration scheme; According to the preliminary resource configuration scheme, resource scheduling optimization and load balancing distribution are performed to obtain a calculation resource configuration. 6.The regional weather prediction method based on a meteorological model according to claim 1, wherein, The weather prediction is performed according to the calculation resource configuration, the updated terrain influence weight matrix and the spatial distribution feature set to obtain a regional weather prediction result, including: The calculation resource configuration is executed, and task execution data is collected in real time; According to the task execution data, the updated terrain influence weight matrix and the spatial distribution feature set, weighted fusion is performed to obtain a fusion feature vector set; According to the fusion feature vector set, prediction calculation and real-time updating are performed to obtain a regional weather prediction result. 7.The regional weather prediction method based on a meteorological model according to claim 1, wherein, According to the regional weather prediction result, a spatial distribution map is generated to obtain a spatial distribution map of local rainfall intensity, and differential correction is performed to obtain a regional weather prediction output, including: Obtain real-time meteorological data; From the weather prediction result, extract the rainfall probability distribution, generate the rainfall intensity index of each grid by grid division, and perform smoothing adjustment on the index exceeding the intensity threshold, combine the real-time meteorological data to generate a spatial distribution map, thereby obtaining a local rainfall intensity spatial distribution map; According to the local rainfall intensity spatial distribution map, the sea-land wind influence factor calculation is performed, and the real-time meteorological data is combined to perform weighted fusion to obtain a sea-land wind modulation parameter; According to the sea-land wind modulation parameter, spatial projection and optimization modulation are performed to output the final regional weather prediction result.

8. A weather prediction system based on a meteorological large model for implementing the method of weather prediction based on a meteorological large model according to any one of claims 1 to 7, characterized in that, Including: The weight matrix acquisition module is used for acquiring a terrain influence weight matrix. The spatial distribution feature module is used for inputting the terrain influence weight matrix into a pre-trained spatial distribution feature model, outputting an optimized spatial distribution feature set, and performing time-space registration to obtain a spatial distribution feature set of local weather. The change trend module is used for extracting a seasonal change trend according to the spatial distribution feature set to obtain a seasonal change trend. The correction coefficient module is configured to perform seasonal deviation correction based on the seasonal change trend to obtain a deviation correction coefficient. The correction coefficient module includes: obtaining real-time environment data and a seasonal feature library; performing smoothing processing on the real-time environment data to obtain a smoothed environment data set; calculating deviation values of the environment data set and the seasonal feature library, and combining the seasonal change trend to perform deviation correction to obtain an initial deviation value set; if an absolute value of any parameter in the initial deviation value set exceeds a corresponding preset deviation threshold, performing offset parameter adjustment to obtain an adjusted offset parameter set, and generating an optimized deviation correction coefficient through linear transformation; wherein the offset parameter set is composed of a horizontal offset angle, a vertical offset angle, and a time difference weight; a center measurement point is constructed from a current measurement point to be corrected, a surrounding measurement point set is constructed, spatial differences between the center measurement point and each surrounding measurement point are calculated, a latitude and longitude point in a WGS-84 coordinate system is used as a geographic coordinate difference, a plane distance in meters is converted from the geographic coordinate difference through a Haversine spherical distance formula, and a distance reciprocal is used as a weight to perform weighted superposition on surrounding measurement point observation values to form the horizontal offset angle; in vertical offset angle calculation, an altitude difference between the center measurement point and the surrounding measurement point is extracted, and a meteorological variable change caused by the altitude is quantitatively corrected according to an empirical climate gradient parameter to compensate for a prediction deviation caused by vertical terrain differences; a time difference weight is obtained by performing cross-correlation analysis on the smoothed time series and the reference time series, finding a time delay corresponding to a correlation coefficient peak, and normalizing the time delay to a weight factor between 0 and 1; The deep coupling module is configured to update a terrain influence weight matrix based on the deviation correction coefficient to obtain an updated terrain influence weight matrix, and perform coupling calculation on the updated terrain influence weight matrix to obtain a coupling parameter set of regional characteristics deep coupling. The resource configuration module is configured to generate a prediction result based on the coupling parameter set, and generate a computing resource configuration if the prediction result is lower than a preset accuracy threshold. The weather prediction module is configured to perform weather prediction based on the computing resource configuration, the updated terrain influence weight matrix, and the spatial distribution feature set to obtain a regional weather prediction result. The prediction output module is configured to generate a spatial distribution map based on the regional weather prediction result to obtain a spatial distribution map of local rainfall intensity, and perform differential correction to obtain a regional weather prediction output.

9. An electronic device, comprising: The processor, the memory, and the computer program stored in the memory and configured to be executed by the processor are included. The processor implements the regional weather prediction method based on the meteorological large model according to any one of claims 1 to 7 when executing the computer program.

Images