Multi-source heterogeneous meteorological data assimilation fusion method

Abstract

The invention discloses a multi-source heterogeneous meteorological data assimilation fusion method, which comprises the following steps that meteorological data of a plurality of data sources are synchronously acquired through a distributed data interface and are preprocessed, and the data sources comprise satellite remote sensing data, ground meteorological station data and meteorological radar data; after finishing preprocessing the meteorological data of the plurality of data sources, fusing the data of the plurality of data sources into a meteorological element value of each meteorological element based on a weighted least square method; and assimilating the fused meteorological element values and a numerical weather forecast model. Compared with the prior art, the method has the following advantages and effects that abnormal data can be accurately obtained and specifically repaired, the weight of a data source is dynamically adjusted through weighting, interference of low-quality data on a fusion result is effectively avoided, and the final prediction accuracy is improved by improving the assimilation step of the numerical weather prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological data processing, and in particular to a multi-source heterogeneous meteorological data assimilation and fusion method. Background Art

[0002] Meteorological data plays a vital role in weather forecasting and climate research. At present, meteorological data mainly comes from different data sources such as ground meteorological stations, satellite remote sensing, meteorological radars, and meteorological soundings. Ground meteorological stations provide local meteorological information with high temporal and spatial resolution, but their coverage is limited and may be affected by equipment failures. Satellite remote sensing can provide global data, especially in terms of atmospheric temperature, humidity, and ocean conditions, but due to its low temporal and spatial resolution and the influence of clouds, it may bring inaccurate data. Meteorological radar is used to monitor meteorological dynamics such as precipitation and wind speed in real time, but its data may be interfered with by terrain and climate, resulting in incomplete or noisy information. Reference can be made to the sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data disclosed in application number CN202411368950.X.

[0003] The differences in spatiotemporal resolution, measurement accuracy, and coverage of the above-mentioned different data sources have caused inconsistencies and errors between meteorological data. In this context, meteorological data assimilation technology came into being. The basic goal of data assimilation is to combine observational data from different data sources with numerical weather forecast models, and by adjusting the initial conditions of the model, make the output of the model closer to the actual meteorological conditions, thereby improving the forecast accuracy. Through data assimilation, not only can the numerical weather forecast model be optimized and the prediction error be corrected in real time, but the forecast results can also be adjusted in time according to new observational data to improve the accuracy of the warning.

[0004] However, the prior art still has the following defects: First, in the process of preprocessing meteorological data from multiple data sources, data missing or outliers often occur due to equipment failure, environmental factors, etc. Existing technologies cannot accurately identify and process these abnormal data to ensure data quality.

[0005] Secondly, in the subsequent fusion and assimilation process, traditional fusion methods are used, such as Kalman filtering, which has limitations in the collaborative optimization of heterogeneous data, resulting in the fusion results being dominated by the error of a single data source. In addition, in the assimilation process with the numerical weather forecast model after fusion, accuracy must also be guaranteed to reduce prediction errors. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-source heterogeneous meteorological data assimilation and fusion method to solve the problems raised in the background technology.

[0007] The above technical objectives of the present invention are achieved through the following technical solutions: A multi-source heterogeneous meteorological data assimilation and fusion method includes the following steps: S100, synchronously acquiring meteorological data from multiple data sources through a distributed data interface and performing preprocessing, the data sources including satellite remote sensing data, ground meteorological station data, and meteorological radar data; S200, after completing the preprocessing of the meteorological data from multiple data sources in S100, based on the weighted least squares method, respectively fuse the data from the multiple data sources into meteorological element values ​​of each meteorological element; S300, assimilating the meteorological element values ​​obtained by fusion in S200 with the numerical weather forecast model; S400, in conjunction with S100, S200 and S300, can dynamically optimize numerical weather forecast model parameters and generate executable weather warning instructions; Among them, S200 is specifically: Multiply the weights of data source type, spatiotemporal proximity, and data quality item by item to generate a comprehensive weight matrix; The data of each meteorological element will correspond to a comprehensive weight matrix, and the fused meteorological element value will be obtained based on the weighted least squares method. By adopting the above technical scheme, the present invention can effectively integrate data from different sources through the assimilation and fusion of multi-source heterogeneous meteorological data, and make up for the shortcomings of a single data source; by fusing multiple data sources through the weighted least squares method, the differences between different data sources can be eliminated, and the fusion results of the same meteorological element can be guaranteed to have higher accuracy and consistency. At the same time, through the weighted fusion of multiple data sources, even if a data source is missing, the data of other data sources can still supplement the relevant information, thereby ensuring the integrity of the final fused data, and finally improving the final prediction accuracy by improving the assimilation step of the numerical weather forecast model.

[0008] It is further configured that the preprocessing process in S100 is: S101, constructing a meteorological element threshold rule base, wherein the meteorological element threshold rule base includes the minimum and maximum values ​​of each meteorological element to preliminarily filter out abnormal data; S102, performing quality inspection on the data points in each meteorological element, and further screening out abnormal data; specifically, adopting a sliding time window mechanism, dynamically inspecting the data points in each meteorological element, and calculating the corresponding local outlier factor according to all the data points in each window based on the window length and step length corresponding to each meteorological element, and if the local outlier factor is greater than a set threshold, all the data points in the window are determined to be abnormal data; Among them, the sliding time window mechanism can automatically adjust the window length and step size according to the volatility of each meteorological element. Specifically, the standard deviation and mean of each data point in the current window of each meteorological element are calculated, and then the calculated standard deviation and mean are divided to obtain the coefficient of variation. When the coefficient of variation is higher than the set maximum threshold, it indicates that the data changes drastically, and the window length is shortened; when the coefficient of variation is lower than the set minimum threshold, it indicates that the data changes slowly, and the window length is extended; The values ​​of the data points at the current moment and the previous moment of each meteorological element are extracted, and divided by the time interval between the current moment and the previous moment to obtain the rate of change at the current moment. When the rate of change at the current moment is greater than the set threshold, the step size is shortened to more accurately capture the trend of rapid changes; when the rate of change at the current moment is less than the set threshold, the step size is extended to reduce the amount of calculation.

[0009] S103, repairing the screened abnormal data; S104: normalize each meteorological element one by one to map different meteorological elements to the same value range.

[0010] By adopting the above technical solution, by building a meteorological element threshold rule library and performing quality detection and repair, these abnormal data can be preliminarily screened out, thereby improving the reliability of the data; then, through quality detection and repair, it is ensured that all data are within the valid numerical range, avoiding the impact of abnormal data on the assimilation process, thereby improving the quality of the final fused data; finally, each meteorological element is normalized so that data from different sources can be compared and fused within a unified numerical range; through the above steps, the data integration process is greatly simplified and the compatibility between data is ensured; Through the sliding time window mechanism, abnormal data caused by external interference or equipment failure can be detected in real time, and the abnormality of the data can be identified by calculating the local outlier factor. It has strong dynamic adaptability and can effectively cope with situations with large data changes.

[0011] The ability to automatically adjust the size of the sliding window according to the volatility of the data helps to more accurately capture rapidly changing meteorological data, such as changes in extreme weather events in a short period of time; this adjustment can flexibly respond to various meteorological changes and improve the accuracy and real-time performance of data processing. Specifically, when meteorological data changes dramatically, the window length and step size can be adjusted by calculating the coefficient of variation and the rate of change, which can more accurately capture rapidly changing meteorological elements, which is of great significance for short-term severe weather events in weather forecasts; when data changes are small or stable, extending the window length and step size can reduce the amount of calculation and improve processing efficiency; this dynamic adjustment mechanism can operate efficiently in different scenarios, avoiding unnecessary computational burdens.

[0012] The further configuration is that S103 is specifically: If the abnormal data is ground meteorological station data; use the Delaunay triangulation algorithm to spatially divide the ground stations, each station is connected to other stations to form a triangle edge, and then determine the adjacency relationship between the station of the abnormal ground meteorological station and the stations of other ground meteorological stations, and then select the stations of the five ground meteorological stations closest to the station of the abnormal ground meteorological station; for the stations of the five selected ground meteorological stations, use a weighted average method to replace the abnormal data, and determine the weight according to the distance between the stations of these ground meteorological stations and the station of the abnormal ground meteorological station, and the weight is inversely proportional to the distance; according to the weights of the stations of the five selected ground meteorological stations, calculate the weighted average to replace the abnormal data; If the abnormal data is satellite remote sensing data or meteorological radar data, the data points within 3 moments before and after the time dimension are used for time interpolation repair.

[0013] By adopting the above technical solution and repairing the ground meteorological station data through the Delaunay triangulation algorithm, the most suitable alternative value can be found in the neighborhood of the abnormal data, thereby ensuring that the repaired data is as close to the true observation value as possible and avoiding the accumulation of errors in the data repair process; by repairing the missing parts of satellite remote sensing data and meteorological radar data through time interpolation, data gaps can be effectively filled, reducing the impact of data missing on the final fusion results, and ensuring the integrity and high quality of the data.

[0014] Further configuration is that S300 is specifically: S301, upgrading the low-resolution data of the numerical weather forecast model to the same resolution as the meteorological data by bicubic interpolation, so that the meteorological data and the data of the numerical weather forecast model can be compared on the same spatial grid; S302: Calculate the difference between the predicted value of the numerical weather forecast model and the meteorological element value.

[0015] By adopting the above technical solution, by assimilating the fused meteorological element values ​​with the numerical weather forecast model, the initial conditions of the model can be corrected and the prediction error can be significantly reduced. Especially in extreme weather forecasting, data assimilation can effectively improve the accuracy of the forecast; the assimilation process can adjust the model prediction according to the actual observation data, reduce the errors introduced by the model assumptions, and enhance the stability and reliability of the forecast results.

[0016] The further configuration is that S302 is specifically: S3021. For the same meteorological element at each grid point, calculate the direct difference between the predicted value and the meteorological element value; S3022, calculating the average value of all grid point difference values ​​to evaluate the systematic deviation of the numerical weather forecast model; then calculating the root mean square error to measure the degree of dispersion between the predicted value and the meteorological element value, reflecting the uncertainty of the numerical weather forecast model; S3023. In combination with the generated comprehensive weight matrix, the difference of each grid point is adjusted according to the weight of the data source to avoid interference of low-quality data in the difference calculation.

[0017] By adopting the above technical solution, by calculating the difference between the predicted value and the meteorological element value, the error of the numerical weather forecast model can be quantified and the deviation of the model prediction can be evaluated; difference analysis and error evaluation can accurately identify which parts of the forecast have deviations, so as to improve the model and forecast algorithm in a targeted manner and further improve the accuracy of the forecast.

[0018] The further configuration is that S3023 also has the following steps: S3024. Visualize the difference results in the form of a spatial heat map, where red indicates a positive deviation of the predicted value that is higher than the meteorological element value, and blue indicates a negative deviation of the predicted value that is lower than the meteorological element value, so as to facilitate intuitive positioning of the deviation concentration area and mark the grid points that exceed the preset threshold.

[0019] By adopting the above technical solution and visualizing the difference results in the form of a spatial heat map, the deviation between the predicted value and the actual observed value can be presented intuitively, helping meteorologists and forecasters to quickly locate areas where deviations are concentrated and optimize subsequent forecasting work.

[0020] Further configuration, S301 is specifically as follows: S3011. Obtain the original grid resolution of the numerical weather forecast model and the actual grid resolution of the meteorological data, and calculate the proportionality coefficient between the two: Based on the original row resolution and original column resolution of the numerical weather forecast model, and the actual row resolution and actual column resolution of the meteorological data, the resolution scaling coefficients are calculated in the row direction and column direction respectively; the row direction scaling coefficient is equal to the row resolution of the meteorological data divided by the row resolution of the original model, and the column direction scaling coefficient is equal to the column resolution of the meteorological data divided by the column resolution of the original model; S3012. Based on the scale factor, dynamically adjust the kernel function parameters of the bicubic interpolation algorithm: According to the larger value of the row-wise proportional coefficient and the column-wise proportional coefficient, the sampling radius of the kernel function is set to 1.5 times the larger value, and the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of the meteorological data; when the density of the meteorological data is higher than the preset threshold, the smoothness coefficient is reduced to retain the detail features; when the density of the meteorological data is lower than the preset threshold, the smoothness coefficient is increased to enhance the data smoothness; S3013. Perform layer-by-layer interpolation processing on the low-resolution data of the numerical weather forecast model: A hierarchical interpolation strategy is used to divide the original low-resolution data into coarse-grained layers and fine-grained layers: The operation of coarse-grained layer interpolation is to perform preliminary interpolation on the original grid based on the bicubic interpolation algorithm to generate intermediate resolution data; The operation of fine-grained layer interpolation is to dynamically adjust the interpolation step size based on the intermediate resolution data according to the local gradient changes of the meteorological data, and finally generate high-resolution grid data consistent with the resolution of the meteorological data; S3014. Verify the spatial coordinate consistency between the interpolated high-resolution grid data and the meteorological data: A preset number of key grid points are randomly selected from the meteorological data as reference points; Calculate the coordinate offset between the interpolated data and the reference point. If the offset exceeds the preset maximum threshold, it is determined as a coordinate deviation. If more than 50% of the reference points have coordinate deviations, the kernel function parameters are readjusted and interpolation is performed; if the deviation ratio is less than 50%, only local corrections are performed on the deviation area; S3015. Align the verified high-resolution grid data with the meteorological data point by point according to spatial coordinates; if there are uncovered areas, use the nearest neighbor interpolation method to supplement the missing values ​​to ensure the integrity and consistency of the fused grid data.

[0021] Further, in S3012, the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of meteorological data. Specifically, if the density of meteorological data changes suddenly in a local area, the smoothness coefficient is adjusted according to the density gradient; when the density gradient is greater than a preset threshold, the smoothness coefficient is reduced to 70% of the original value to suppress excessive interpolation smoothing; when the density gradient is less than the preset threshold, the smoothness coefficient is increased to 130% of the original value to eliminate data noise interference; In the above-mentioned S3013, when performing the fine-grained layer interpolation operation, if the change trends of the meteorological elements of adjacent grids are consistent, linear interpolation is adopted; if the change trends are significantly different, a nonlinear interpolation algorithm is switched to avoid interpolation distortion.

[0022] Further settings, S400 specifically: S401, perform source tracing analysis on the marked grid points that exceed the preset threshold, extract the corresponding data source type, spatiotemporal resolution characteristics and abnormal fluctuation records of the comprehensive weight matrix, and generate an abnormal source tracing report; S402, based on the anomaly tracing report, dynamically adjust the empirical parameter weights according to the deviation direction and magnitude for the meteorological element model parameters associated with the marked grid point in the numerical weather forecast model, including the following sub-steps: S4021, extracting the corresponding empirical parameter set from the numerical weather forecast model according to the meteorological element type of the marked grid point, and associating the predicted value calculated in S3021 with the meteorological element value after fusion in S200, and establishing a mapping relationship table between parameters and deviations; S4022, determining the deviation direction based on the direct difference between the predicted value and the meteorological element value at the same grid point in S3021; ​​specifically: S4021.1. Obtain the direct difference between the predicted value and the meteorological element value at the same grid point in S3021, and detect the sign consistency of the difference in the continuous time series based on the preset time window; if more than 80% of the differences are positive values ​​in the same window, a positive deviation determination is triggered; if more than 80% of the differences are negative values, a negative deviation determination is triggered; S4021.2. If a positive deviation judgment is triggered, it is confirmed that the predicted value of the numerical weather forecast model is continuously higher than the fused meteorological element value, and a gain parameter weight reduction instruction is generated; if a negative deviation judgment is triggered, it is confirmed that the predicted value is continuously lower than the meteorological element value, and a compensation parameter weight increase instruction is generated; S4021.3. According to the confirmation result of the deviation direction in S4021.2, add dynamic marks to the associated meteorological element model parameters in the model parameter correction coefficient table. The mark types include weight attenuation factors or weight enhancement factors. The weight attenuation factors and weight enhancement factors are proportional to the average value of all grid point difference values ​​calculated in S3022. S4021.4. Execute regional association expansion. If the same meteorological element triggers the same deviation direction judgment at three or more adjacent grid points, the association parameter weight adjustment instruction is regionalized and the single-point adjustment amount is expanded to the regional range. The adjustment amplitude decays with the grid distance. S4023, based on the average value of all grid point difference values ​​calculated in S3022, combined with the preset error reference value, calculate the parameter weight adjustment amount: If the predicted value is higher than the meteorological element value, the directional coefficient is assigned a negative value and the corresponding parameter weight is proportionally reduced; if the predicted value is lower than the meteorological element value, the directional coefficient is assigned a positive value and the corresponding parameter weight is proportionally increased; Divide the average value by the preset error reference value to obtain a normalized adjustment ratio; Multiply the directional coefficient by the normalized adjustment ratio to obtain the final adjustment of the parameter weight, ensuring that the adjustment amplitude is linearly related to the degree of deviation; S4024, perform smoothing constraints to limit the adjustment range of a single parameter weight to no more than 10% of the original weight. If the calculated value exceeds the range, it is truncated according to the threshold to ensure model stability; S403, use the historical assimilated data to train a lightweight neural network, input the adjusted difference in S3023 into the network, predict the short-term meteorological element change trend, and feed back the prediction results to the weight allocation module of S200 to update the comprehensive weight matrix.

[0023] Further, the S403 is specifically as follows: S4031, extract the difference adjusted by S3023, the generated comprehensive weight matrix and the corresponding timestamp from the historical assimilation data, and generate a time series training sample with a label according to the sliding time window mechanism, where the label is the actual change value of the meteorological element in the next 1 hour; S4032. Construct a lightweight neural network, including an input layer, two convolutional layers, a single-layer long short-term memory unit, and a fully connected output layer, where the convolution kernel size is set to 3×3 and the step size matches the meteorological grid resolution; S4033, using an adaptive learning rate mechanism, with the initial learning rate set to 0.001. If the validation set loss does not decrease for three consecutive training cycles, the learning rate is decayed to 50% of the original value. At the same time, an early stopping mechanism is introduced to prevent overfitting. S4034, input the adjusted difference value of S3023 at the current moment into the trained network, output the forecast value of the meteorological element change trend in the next hour, and combine the forecast value with the real-time stability index of the data source in S200 to generate a weight correction factor; S4035. Based on the weight correction factors generated in S4034, the comprehensive weight matrix is ​​dynamically adjusted. Specifically, the weights of data sources with high consistency between the forecast trend and the real-time data are increased, and the weights of data sources with low consistency are decreased. The updated weight matrix is ​​directly applied to the next data fusion.

[0024] In summary, the present invention has the following beneficial effects: The present invention can accurately identify abnormal data and carry out targeted repairs, and dynamically adjust the weight of the data source through weighting, effectively avoiding the interference of low-quality data on the fusion result, and then improve the final prediction accuracy by improving the assimilation step of the numerical weather forecast model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the process of the embodiment. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0027] As attached Figure 1 As shown; This embodiment discloses a method for assimilating and fusing multi-source heterogeneous meteorological data, including the following steps: S100, synchronously acquiring meteorological data from multiple data sources through a distributed data interface and performing preprocessing, the data sources including satellite remote sensing data, ground meteorological station data, and meteorological radar data; S200, after completing the preprocessing of the meteorological data from multiple data sources in S100, based on the weighted least squares method, respectively fuse the data from the multiple data sources into meteorological element values ​​of each meteorological element; S300, assimilating the meteorological element values ​​obtained by fusion in S200 with the numerical weather forecast model.

[0028] S400, in conjunction with S100, S200 and S300, can dynamically optimize numerical weather forecast model parameters and generate executable meteorological warning instructions.

[0029] Among them, the pre-processing process in S100 is: S101. Construct a meteorological element threshold rule base, which includes the minimum and maximum values ​​of each meteorological element to preliminarily screen out abnormal data; define the effective range of temperature as -90℃ to 60℃, which covers common climate conditions from extreme cold to tropical high temperatures; set the effective range of air pressure to 500hPa to 1100hPa to adapt to changes in air pressure at different altitudes; define the effective range of relative humidity to be 0% to 100%, representing humidity conditions from completely dry to saturated; set the effective range of wind speed to 0m / s to 75m / s to adapt to regular weather conditions and special meteorological events, such as typhoons or tornadoes.

[0030] S102, performing quality inspection on the data points in each meteorological element, and further screening out abnormal data; S103, repairing the screened abnormal data; S104, normalizing each meteorological element one by one to map different meteorological elements to the same numerical range; the temperature is normalized by Min-Max, the air pressure is logarithmically transformed and then normalized by Min-Max, the humidity is normalized by linear scaling, and the wind speed is first de-skewed and then normalized by Min-Max.

[0031] Among them, S102 is specifically: A sliding time window mechanism is adopted to dynamically detect the data points in each meteorological element, and based on the window length and step size corresponding to each meteorological element, the corresponding local outlier factor is calculated according to all the data points in each window. If the local outlier factor is greater than the set threshold of 0.5, all data points in the window are judged as abnormal data.

[0032] Among them, the sliding time window mechanism can automatically adjust the window length and step size according to the volatility of each meteorological element; specifically: Calculate the standard deviation and mean of each data point in the current window of each meteorological element, and then divide the calculated standard deviation and mean to get the coefficient of variation. When the coefficient of variation is higher than the set maximum threshold, it indicates that the data changes dramatically, and the window length is shortened; when the coefficient of variation is lower than the set minimum threshold, it indicates that the data changes slowly, and the window length is extended.

[0033] The formula for calculating standard deviation is: ; Among them, σ t is the standard deviation, x i is the value of the i-th data point in the current sliding window, μ is the mean of the data points in the current window, and N represents the number of data points contained in the current window; The coefficient of variation calculation formula is: CV t =σ t / μ; Among them, CV t represents the coefficient of variation; The formula for dynamically adjusting the window length is defined as: T t+1 =max(T min ,min(T max ,T t +α(β-CV t ))); Among them, T t+1 is the new window length, T t is the current window length, T min and T max The minimum and maximum window lengths are 15 minutes and 60 minutes, respectively. α is the adjustment rate and β is the expected value. Based on the calculated standard deviation and coefficient of variation of each data point in the current window, the coefficient of variation is compared with the set maximum threshold of 0.1 and minimum threshold of 0.05; If CV t >0.1, the window length is reduced to T t+1 =T t -α(CV t -0.1); If CV t <0.05, the window length increases to T t+1 =T t +α(0.05-CV t ).

[0034] The values ​​of the data points at the current moment and the previous moment of each meteorological element are extracted, and divided by the time interval between the current moment and the previous moment to obtain the rate of change at the current moment. When the rate of change at the current moment is greater than the set threshold, the step size is shortened to more accurately capture the trend of rapid changes; when the rate of change at the current moment is less than the set threshold, the step size is extended to reduce the amount of calculation.

[0035] The formula for calculating the rate of change is: ; Where Δx t represents the rate of change at the current time t, x t represents the value of the data point at the current time t, x t-1 represents the value of the data point at the previous time t-1, and Δt represents the time interval; The formula for dynamically adjusting the step size is defined as: Δt t+1 =max(Δt min ,min(Δt max ,Δt t +γ(δ-Δx t ))); Among them, Δt t+1 is the step size of the next calculation, Δt min and Δt max The minimum step size is 1 minute and the maximum step size is 15 minutes, γ is the adjustment rate, and δ is the desired change rate threshold; If Δx t >0.1, the step length is shortened to Δt t+1 =Δt t -γ(Δx t -0.1); If Δx t <0.05, step length extended to Δt t+1 =Δt t +γ(0.05-Δx t ).

[0036] Among them, S103 is specifically: If the abnormal data is ground meteorological station data; use the Delaunay triangulation algorithm to spatially divide the ground stations, each station is connected to other stations to form a triangle edge, and then determine the adjacency relationship between the station of the abnormal ground meteorological station and the stations of other ground meteorological stations, and then select the stations of the five ground meteorological stations closest to the station of the abnormal ground meteorological station; for the stations of the five selected ground meteorological stations, use a weighted average method to replace the abnormal data, and determine the weight according to the distance between these ground meteorological stations and the station of the abnormal ground meteorological station, and the weight is inversely proportional to the distance; according to the weights of the five selected ground meteorological stations, calculate the weighted average to replace the abnormal data.

[0037] The specific process is shown in the following formula: ; in, is the abnormal data after interpolation, y j is the jth ground weather station data, w j is the site weight of the jth ground weather station.

[0038] If the abnormal data is satellite remote sensing data or meteorological radar data, the data points within 3 moments before and after the time dimension are used for time interpolation repair.

[0039] Assuming the current time is t, select the data points at time t-3, t-2, t-1, t+1, t+2, and t+3.

[0040] Among them, S200 is specifically: S201. Due to the significant differences in the temporal and spatial resolution and error characteristics of satellite remote sensing, ground meteorological stations, and meteorological radar data, dynamic weights need to be assigned to each data source and meteorological element to balance their contributions. S202, multiplying the weights of data source type, spatiotemporal proximity, and data quality item by item to generate a comprehensive weight matrix; Comprehensive weight matrix W k The formula is: W k =W source ⊙W time ⊙Q quality； Among them, W source Represents the data source type weight matrix, W time represents the spatiotemporal proximity weight matrix, W space represents the data quality weight matrix, and ⊙ represents element-wise multiplication.

[0041] S203, based on S201 and S202, each meteorological element data will correspond to a comprehensive weight matrix, and the fused meteorological element value is obtained based on the weighted least square method; The process is as follows: ; in, is the fused meteorological element value, a i is the data of the ith data source, is the weight of the i-th data source.

[0042] Among them, S201 is specifically: Assign weights based on data source type; For ground weather station data, the highest basic weight of 0.6 is assigned in the corresponding geographical location; For satellite remote sensing data, the weight of wide-area coverage areas is increased to 0.5, and the weight of ground station dense areas is reduced to 0.3 to avoid conflicts with ground meteorological station data; For meteorological radar data, the weight is dynamically adjusted according to scene changes; for example, in the case of sudden changes in precipitation and wind speed caused by short-term severe weather changes, the weight is increased to 0.7.

[0043] Assign weights based on spatiotemporal proximity; dynamically adjust weights based on target fusion time according to time difference; weights for data within ±15 minutes of target fusion time are 1.0, and decay to 0.3 beyond 1 hour; The weights are assigned based on the data quality. For the abnormal data repaired in S103, the weight is reduced proportionally according to the confidence of the repair algorithm. The confidence of the Delaunay triangulation algorithm is 0.8, and the confidence of the time interpolation is 0.6. If a data point is judged as abnormal data multiple times, specifically, the local outlier factor in S102 is > 0.5 for more than 3 times, it is directly assigned a weight of 0 to avoid contaminating the fusion result.

[0044] Among them, S300 is specifically: S301, upgrading the low-resolution data of the numerical weather forecast model to the same resolution as the meteorological data by bicubic interpolation, so that the meteorological data and the data of the numerical weather forecast model can be compared on the same spatial grid; S302: Calculate the difference between the predicted value of the numerical weather forecast model and the meteorological element value.

[0045] Among them, S302 is specifically: S3021. For the same meteorological element at each grid point, calculate the direct difference between the predicted value and the meteorological element value; ; Among them, ΔE(x,y,t) is the deviation between the predicted value and the meteorological element value, E NWP is the predicted value of the numerical weather prediction model, is the specific value of the meteorological element at the spatial coordinate (x, y) and time t; The above formula can directly reflect the prediction deviation of a local area, for example, the temperature in a certain area is overestimated by 2°C.

[0046] S3022, calculating the average value of all grid point difference values ​​to evaluate the systematic deviation of the numerical weather forecast model; then calculating the root mean square error to measure the degree of dispersion between the predicted value and the meteorological element value, reflecting the uncertainty of the numerical weather forecast model; The process is as follows: ; in, represents the average value of all grid point differences, and N represents the total number of all spatial grid points; ; Where RMSE represents the root mean square error, and N represents the total number of all spatial grid points.

[0047] S3023. In combination with the comprehensive weight matrix generated in S202, the difference of each grid point is adjusted according to the weight of the data source to avoid interference of low-quality data on the difference calculation.

[0048] ; Among them, W k is the comprehensive weight matrix generated in S202, corresponding to the spatial coordinates (x, y) and time t; Among them, after S3023, the following steps are also required: S3024. Visualize the difference results in the form of a spatial heat map, where red indicates a positive deviation of the predicted value that is higher than the meteorological element value, and blue indicates a negative deviation of the predicted value that is lower than the meteorological element value, so as to facilitate intuitive positioning of the deviation concentration area and mark the grid points that exceed the preset threshold.

[0049] Among them, S301 is specifically: S3011. Obtain the original grid resolution of the numerical weather forecast model and the actual grid resolution of the meteorological data, and calculate the proportionality coefficient between the two: Based on the original row resolution and original column resolution of the numerical weather forecast model, and the actual row resolution and actual column resolution of the meteorological data, the resolution scaling coefficients are calculated in the row direction and column direction respectively; the row direction scaling coefficient is equal to the row resolution of the meteorological data divided by the row resolution of the original model, and the column direction scaling coefficient is equal to the column resolution of the meteorological data divided by the column resolution of the original model; S3012. Based on the scale factor, dynamically adjust the kernel function parameters of the bicubic interpolation algorithm: According to the larger value of the row-wise proportional coefficient and the column-wise proportional coefficient, the sampling radius of the kernel function is set to 1.5 times the larger value, and the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of the meteorological data; when the density of the meteorological data is higher than the preset threshold, the smoothness coefficient is reduced to retain the detail features; when the density of the meteorological data is lower than the preset threshold, the smoothness coefficient is increased to enhance the data smoothness; S3013. Perform layer-by-layer interpolation processing on the low-resolution data of the numerical weather forecast model: A hierarchical interpolation strategy is used to divide the original low-resolution data into coarse-grained layers and fine-grained layers: The operation of coarse-grained layer interpolation is to perform preliminary interpolation on the original grid based on the bicubic interpolation algorithm to generate intermediate resolution data; The operation of fine-grained layer interpolation is to dynamically adjust the interpolation step size based on the intermediate resolution data according to the local gradient changes of the meteorological data, and finally generate high-resolution grid data consistent with the resolution of the meteorological data; S3014. Verify the spatial coordinate consistency between the interpolated high-resolution grid data and the meteorological data: A preset number of key grid points are randomly selected from the meteorological data as reference points; Calculate the coordinate offset between the interpolated data and the reference point. If the offset exceeds the preset maximum threshold, it is determined as a coordinate deviation. If more than 50% of the reference points have coordinate deviations, the kernel function parameters are readjusted and interpolation is performed; if the deviation ratio is less than 50%, only local corrections are performed on the deviation area; S3015. Align the verified high-resolution grid data with the meteorological data point by point according to spatial coordinates; if there are uncovered areas, use the nearest neighbor interpolation method to supplement the missing values ​​to ensure the integrity and consistency of the fused grid data.

[0050] Among them, in S3012, the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of meteorological data: if the density of meteorological data changes suddenly in a local area, the smoothness coefficient is adjusted according to the density gradient; when the density gradient is greater than a preset threshold, the smoothness coefficient is reduced to 70% of the original value to suppress excessive interpolation smoothing; when the density gradient is less than the preset threshold, the smoothness coefficient is increased to 130% of the original value to eliminate data noise interference; In the above-mentioned S3013, when performing the fine-grained layer interpolation operation, if the change trends of the meteorological elements of adjacent grids are consistent, linear interpolation is adopted; if the change trends are significantly different, a nonlinear interpolation algorithm is switched to avoid interpolation distortion.

[0051] Among them, S400 is specifically: S401, perform source tracing analysis on the marked grid points that exceed the preset threshold, extract the corresponding data source type, spatiotemporal resolution characteristics and abnormal fluctuation records of the comprehensive weight matrix, and generate an abnormal source tracing report; S402, based on the anomaly tracing report, dynamically adjust the empirical parameter weights according to the deviation direction and magnitude for the meteorological element model parameters associated with the marked grid point in the numerical weather forecast model, including the following sub-steps: S4021, extracting the corresponding empirical parameter set from the numerical weather forecast model according to the meteorological element type of the marked grid point, and associating the predicted value calculated in S3021 with the meteorological element value after fusion in S200, and establishing a mapping relationship table between parameters and deviations; S4022, determining the deviation direction based on the direct difference between the predicted value and the meteorological element value at the same grid point in S3021; ​​specifically: S4021.1. Obtain the direct difference between the predicted value and the meteorological element value at the same grid point in S3021, and detect the sign consistency of the difference in the continuous time series based on the preset time window; if more than 80% of the differences are positive values ​​in the same window, a positive deviation determination is triggered; if more than 80% of the differences are negative values, a negative deviation determination is triggered; S4021.2. If a positive deviation judgment is triggered, it is confirmed that the predicted value of the numerical weather forecast model is continuously higher than the fused meteorological element value, and a gain parameter weight reduction instruction is generated; if a negative deviation judgment is triggered, it is confirmed that the predicted value is continuously lower than the meteorological element value, and a compensation parameter weight increase instruction is generated; S4021.3. According to the confirmation result of the deviation direction in S4021.2, add dynamic marks to the associated meteorological element model parameters in the model parameter correction coefficient table. The mark types include weight attenuation factors or weight enhancement factors. The weight attenuation factors and weight enhancement factors are proportional to the average value of all grid point difference values ​​calculated in S3022. S4021.4. Execute regional association expansion. If the same meteorological element triggers the same deviation direction judgment at three or more adjacent grid points, the association parameter weight adjustment instruction is regionalized and the single-point adjustment amount is expanded to the regional range. The adjustment amplitude decays with the grid distance. S4023, based on the average value of all grid point difference values ​​calculated in S3022, combined with the preset error reference value, calculate the parameter weight adjustment amount: If the predicted value is higher than the meteorological element value, the directional coefficient is assigned a negative value and the corresponding parameter weight is proportionally reduced; if the predicted value is lower than the meteorological element value, the directional coefficient is assigned a positive value and the corresponding parameter weight is proportionally increased; Divide the average value by the preset error reference value to obtain a normalized adjustment ratio; Multiply the directional coefficient by the normalized adjustment ratio to obtain the final adjustment of the parameter weight, ensuring that the adjustment amplitude is linearly related to the degree of deviation; S4024, perform smoothing constraints to limit the adjustment range of a single parameter weight to no more than 10% of the original weight. If the calculated value exceeds the range, it is truncated according to the threshold to ensure model stability; S403, use the historical assimilated data to train a lightweight neural network, input the adjusted difference in S3023 into the network, predict the short-term meteorological element change trend, and feed back the prediction results to the weight allocation module in S200 to update the comprehensive weight matrix in S202.

[0052] Among them, S403 is specifically: S4031, extract the difference adjusted by S3023, the comprehensive weight matrix generated by S202 and the corresponding timestamp from the historical assimilated data, and generate a time series training sample with a label according to the sliding time window mechanism, where the label is the actual change value of the meteorological element in the next 1 hour; S4032. Construct a lightweight neural network, including an input layer, two convolutional layers, a single-layer long short-term memory unit, and a fully connected output layer, where the convolution kernel size is set to 3×3 and the step size matches the meteorological grid resolution; S4033, using an adaptive learning rate mechanism, with the initial learning rate set to 0.001. If the validation set loss does not decrease for three consecutive training cycles, the learning rate is decayed to 50% of the original value. At the same time, an early stopping mechanism is introduced to prevent overfitting. S4034, input the adjusted difference value of S3023 at the current moment into the trained network, output the forecast value of the meteorological element change trend in the next hour, and combine the forecast value with the real-time stability index of the data source in S200 to generate a weight correction factor; S4035. Based on the weight correction factor generated in S4034, the comprehensive weight matrix in S202 is dynamically adjusted, specifically: the weight of the data source with high consistency between the predicted trend and the real-time data is increased, and the weight of the data source with low consistency is reduced. The updated weight matrix is ​​directly applied to the next data fusion.

[0053] Application Example 1 Generate a virtual typhoon based on historical typhoon parameters, with the maximum wind speed in the center simulated to be 60m / s, and the path simulated to move toward the southeast coast of China; use WRF to generate simulated satellite, ground station, and radar data; Abnormal data is injected into S100 to simulate real interference, specifically: At a certain ground weather station, the instantaneous wind speed reached 80m / s; A wind speed of 70 m / s is injected into the outer area of ​​the typhoon, which is in line with the effective range of wind speed of 0 m / s to 75 m / s in constructing the meteorological element threshold rule base, but this value deviates significantly from the average wind speed of 50 m / s at the surrounding stations in the local window.

[0054] The wind speed at a ground weather station increased from 50m / s to 70m / s within 10 minutes. , greater than the threshold 0.1, to trigger dynamic step size adjustment.

[0055] In S101, the instantaneous wind speed of a ground meteorological station reaches 80 m / s, which obviously exceeds the threshold value of 75 m / s and is determined to be abnormal data; the wind speed data of 70 m / s does not exceed the threshold value of 75 m / s, but needs to be further detected through S102.

[0056] In S102, the coefficient of variation CV t =0.15, which is greater than the maximum threshold of 0.1, and the window length is shortened from 30 minutes to 15 minutes; in the shortened window, the local outlier factor LOF of the wind speed of 70m / s is 0.62, which is greater than the set threshold of 0.5 and is judged as abnormal data.

[0057] In S103, the abnormal data in the ground meteorological station are selected from five adjacent stations through the Delaunay triangulation algorithm, and the weighted average repair value is 53m / s, and the error is reduced from 70m / s to 53m / s. Compared with the simulated true value of 55m / s, the error is <4%.

[0058] In S200, the weights are distributed as follows: in the typhoon core area, i.e. within the simulation radius of 100 km, the radar weight is increased to 0.7 and the satellite weight is reduced to 0.3. Compared with the actual radar weight of 0.72 and the satellite weight of 0.28 in the simulation, the error is <5%.

[0059] The fusion accuracy comparison is as follows: The root mean square error of wind speed from single radar data is 6.8m / s, and the fused meteorological element value The root mean square error of wind speed is 3.2m / s; the accuracy of wind speed prediction has been significantly improved.

[0060] After assimilation in S300, the typhoon's 24-hour path error was reduced from 65km to 28km. By assimilating the fused meteorological element values ​​into the numerical weather forecast model, the accuracy of the model was significantly improved and the forecast error was reduced.

[0061] Application Example 2 Simulate a cold wave process, covering the mid-latitude area, with an average temperature drop of 15°C within 24 hours, local wind speed reaching 20m / s, and accompanied by large-scale heavy snow. Use WRF to generate simulated data and inject data anomalies.

[0062] In S100, a ground meteorological station reported an extremely low temperature due to a sensor failure, which exceeded the actual value by 12, triggering a dynamic detection mechanism; in addition, the temperature of a ground meteorological station dropped sharply within 10 minutes, and the rate of change exceeded the threshold of 0.1, triggering a dynamic adjustment of the step size; the reflectivity factor of a radar in the meteorological radar increased abnormally, exceeding the actual value by 20dBZ; in satellite remote sensing, the surface temperature data was missing due to cloud cover in some local areas.

[0063] In S102, the temperature factor abnormality is measured by the coefficient of variation CV t =0.077 and local outlier factor LOF=0.62 were used to identify the abnormal ground meteorological stations, and they were repaired based on the Delaunay triangulation algorithm with an error of ≤0.5℃. Time interpolation was used to repair radar data, and the error was reduced to within 3dBZ after repair. Time interpolation was used to fill the missing areas in satellite data repair, and the root mean square error was ≤1.2℃.

[0064] In S200, the basic weight of the ground meteorological station data of the temperature element is 0.6, the weight of the meteorological radar data is 0.378, and the weight of the satellite remote sensing data is 0.28. The temperature fusion error is reduced from 5°C of a single data source to 0.7°C.

[0065] In S300, the resolution of the numerical weather forecast model was increased from 10km to 1km, achieving grid-level alignment; after deviation correction, the root mean square error of temperature prediction was reduced from 4.2℃ to 1.8℃, an improvement of 57%.

[0066] At the same time, the heat map shows a positive deviation in the coastal area, which is 2°C higher than the forecast, and is corrected to within ±0.5°C after assimilation.

[0067] In the second application example, the temperature accuracy of the cold wave process can be improved by more than 50%, and the ability to capture extreme weather is significantly enhanced, verifying the effectiveness of dynamic fusion and assimilation of multi-source data.

[0068] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A multi-source heterogeneous meteorological data assimilation and fusion method, characterized in that: The following steps are included: S100, synchronously acquiring meteorological data from multiple data sources through a distributed data interface and performing preprocessing, the data sources including satellite remote sensing data, ground meteorological station data, and meteorological radar data; S200, after completing the preprocessing of the meteorological data from multiple data sources in S100, based on the weighted least squares method, respectively fuse the data from the multiple data sources into meteorological element values ​​of each meteorological element; S300, assimilating the meteorological element values ​​obtained by fusion in S200 with the numerical weather forecast model; S400, in conjunction with S100, S200 and S300, can dynamically optimize numerical weather forecast model parameters and generate executable weather warning instructions; Among them, S200 is specifically: Multiply the weights of data source type, spatiotemporal proximity, and data quality item by item to generate a comprehensive weight matrix; The data of each meteorological element will correspond to a comprehensive weight matrix, and the fused meteorological element value will be obtained based on the weighted least squares method.

2. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 1 is characterized by: The pre-processing process in S100 is as follows: S101, constructing a meteorological element threshold rule base, wherein the meteorological element threshold rule base includes the minimum and maximum values ​​of each meteorological element to preliminarily filter out abnormal data; S102, performing quality inspection on the data points in each meteorological element, and further screening out abnormal data; specifically, adopting a sliding time window mechanism, dynamically inspecting the data points in each meteorological element, and calculating the corresponding local outlier factor according to all the data points in each window based on the window length and step length corresponding to each meteorological element, and if the local outlier factor is greater than a set threshold, all the data points in the window are determined to be abnormal data; Among them, the sliding time window mechanism can automatically adjust the window length and step size according to the volatility of each meteorological element. Specifically, the standard deviation and mean of each data point in the current window of each meteorological element are calculated, and then the calculated standard deviation and mean are divided to obtain the coefficient of variation. When the coefficient of variation is higher than the set maximum threshold, it indicates that the data changes drastically, and the window length is shortened; when the coefficient of variation is lower than the set minimum threshold, it indicates that the data changes slowly, and the window length is extended; Extract the values ​​of the data points at the current moment and the previous moment of each meteorological element, and divide them by the time interval between the current moment and the previous moment to get the change rate at the current moment. When the change rate at the current moment is greater than the set threshold, shorten the step length to more accurately capture the trend of rapid changes; when the change rate at the current moment is less than the set threshold, extend the step length to reduce the amount of calculation; S103, repairing the screened abnormal data; S104: normalize each meteorological element one by one to map different meteorological elements to the same value range.

3. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 2 is characterized by: S103 is specifically: If the abnormal data is ground meteorological station data; use the Delaunay triangulation algorithm to spatially divide the ground stations, each station is connected to other stations to form a triangle edge, and then determine the adjacency relationship between the station of the abnormal ground meteorological station and the stations of other ground meteorological stations, and then select the stations of the five ground meteorological stations closest to the station of the abnormal ground meteorological station; for the stations of the five selected ground meteorological stations, use a weighted average method to replace the abnormal data, and determine the weight according to the distance between the stations of these ground meteorological stations and the station of the abnormal ground meteorological station, and the weight is inversely proportional to the distance; according to the weights of the stations of the five selected ground meteorological stations, calculate the weighted average to replace the abnormal data; If the abnormal data is satellite remote sensing data or meteorological radar data, the data points within 3 moments before and after the time dimension are used for time interpolation repair.

4. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 2 is characterized by: S300 is specifically: S301, upgrading the low-resolution data of the numerical weather forecast model to the same resolution as the meteorological data by bicubic interpolation, so that the meteorological data and the data of the numerical weather forecast model can be compared on the same spatial grid; S302, calculating the difference between the predicted value of the numerical weather forecast model and the meteorological element value; specifically: S3021. For the same meteorological element at each grid point, calculate the direct difference between the predicted value and the meteorological element value; S3022, calculating the average value of all grid point difference values ​​to evaluate the systematic deviation of the numerical weather forecast model; then calculating the root mean square error to measure the degree of dispersion between the predicted value and the meteorological element value, reflecting the uncertainty of the numerical weather forecast model; S3023. In combination with the generated comprehensive weight matrix, the difference value of each grid point is adjusted according to the weight of the data source to avoid interference of low-quality data on the difference calculation; S3024. Visualize the difference results in the form of a spatial heat map, where red indicates a positive deviation of the predicted value that is higher than the meteorological element value, and blue indicates a negative deviation of the predicted value that is lower than the meteorological element value, so as to facilitate intuitive positioning of the deviation concentration area and mark the grid points that exceed the preset threshold.

5. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 4 is characterized by: S301 is as follows: S3011. Obtain the original grid resolution of the numerical weather forecast model and the actual grid resolution of the meteorological data, and calculate the proportionality coefficient between the two: Based on the original row resolution and original column resolution of the numerical weather forecast model and the actual row resolution and actual column resolution of the meteorological data, the resolution scaling coefficient is calculated in the row direction and the column direction respectively; The row-wise scaling factor is equal to the row resolution of the meteorological data divided by the row resolution of the original model, and the column-wise scaling factor is equal to the column resolution of the meteorological data divided by the column resolution of the original model; S3012. Based on the scale factor, dynamically adjust the kernel function parameters of the bicubic interpolation algorithm: According to the larger value of the row-wise proportional coefficient and the column-wise proportional coefficient, the sampling radius of the kernel function is set to 1.5 times the larger value, and the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of the meteorological data; when the density of the meteorological data is higher than the preset threshold, the smoothness coefficient is reduced to retain the detail features; when the density of the meteorological data is lower than the preset threshold, the smoothness coefficient is increased to enhance the data smoothness; S3013. Perform layer-by-layer interpolation processing on the low-resolution data of the numerical weather forecast model: A hierarchical interpolation strategy is used to divide the original low-resolution data into coarse-grained layers and fine-grained layers: The operation of coarse-grained layer interpolation is to perform preliminary interpolation on the original grid based on the bicubic interpolation algorithm to generate intermediate resolution data; The operation of fine-grained layer interpolation is to dynamically adjust the interpolation step size based on the intermediate resolution data according to the local gradient changes of the meteorological data, and finally generate high-resolution grid data consistent with the resolution of the meteorological data; S3014. Verify the spatial coordinate consistency between the interpolated high-resolution grid data and the meteorological data: A preset number of key grid points are randomly selected from the meteorological data as reference points; Calculate the coordinate offset between the interpolated data and the reference point. If the offset exceeds the preset maximum threshold, it is determined as a coordinate deviation. If more than 50% of the reference points have coordinate deviations, the kernel function parameters are readjusted and interpolation is performed; if the deviation ratio is less than 50%, only local corrections are performed on the deviation area; S3015. Align the verified high-resolution grid data with the meteorological data point by point according to spatial coordinates; if there are uncovered areas, use the nearest neighbor interpolation method to supplement the missing values ​​to ensure the integrity and consistency of the fused grid data.

6. A multi-source heterogeneous meteorological data assimilation and fusion method according to claim 5, characterized in that: In S3012, the smoothness coefficient is adaptively adjusted in combination with the spatial distribution density of meteorological data. Specifically, if the density of meteorological data changes suddenly in a local area, the smoothness coefficient is adjusted according to the density gradient. When the density gradient is greater than a preset threshold, the smoothness coefficient is reduced to 70% of the original value to suppress excessive interpolation smoothing. When the density gradient is less than the preset threshold, the smoothness coefficient is increased to 130% of the original value to eliminate data noise interference. In the above-mentioned S3013, when performing the fine-grained layer interpolation operation, if the change trends of the meteorological elements of adjacent grids are consistent, linear interpolation is adopted; if the change trends are significantly different, a nonlinear interpolation algorithm is switched to avoid interpolation distortion.

7. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 4 is characterized by: S400 is specifically: S401, perform source tracing analysis on the marked grid points that exceed the preset threshold, extract the corresponding data source type, spatiotemporal resolution characteristics and abnormal fluctuation records of the comprehensive weight matrix, and generate an abnormal source tracing report; S402, based on the anomaly tracing report, dynamically adjust the empirical parameter weights according to the deviation direction and magnitude for the meteorological element model parameters associated with the marked grid point in the numerical weather forecast model, including the following sub-steps: S4021, extracting the corresponding empirical parameter set from the numerical weather forecast model according to the meteorological element type of the marked grid point, and associating the predicted value calculated in S3021 with the meteorological element value after fusion in S200, and establishing a mapping relationship table between parameters and deviations; S4022, determining the deviation direction based on the direct difference between the predicted value and the meteorological element value at the same grid point in S3021; ​​specifically: S4021.

1. Obtain the direct difference between the predicted value and the meteorological element value at the same grid point in S3021, and detect the sign consistency of the difference in the continuous time series based on the preset time window; if more than 80% of the differences are positive values ​​in the same window, a positive deviation determination is triggered; if more than 80% of the differences are negative values, a negative deviation determination is triggered; S4021.

2. If a positive deviation judgment is triggered, it is confirmed that the predicted value of the numerical weather forecast model is continuously higher than the fused meteorological element value, and a gain parameter weight reduction instruction is generated; if a negative deviation judgment is triggered, it is confirmed that the predicted value is continuously lower than the meteorological element value, and a compensation parameter weight increase instruction is generated; S4021.

3. According to the confirmation result of the deviation direction in S4021.2, add dynamic marks to the associated meteorological element model parameters in the model parameter correction coefficient table. The mark types include weight attenuation factors or weight enhancement factors. The weight attenuation factors and weight enhancement factors are proportional to the average value of all grid point difference values ​​calculated in S3022. S4021.

4. Execute regional association expansion. If the same meteorological element triggers the same deviation direction judgment at three or more adjacent grid points, the association parameter weight adjustment instruction is regionalized and the single-point adjustment amount is expanded to the regional range. The adjustment amplitude decays with the grid distance. S4023, based on the average value of all grid point difference values ​​calculated in S3022, combined with the preset error reference value, calculate the parameter weight adjustment amount: If the predicted value is higher than the meteorological element value, the directional coefficient is assigned a negative value and the corresponding parameter weight is reduced proportionally; If the predicted value is lower than the meteorological element value, the directional coefficient is assigned a positive value and the corresponding parameter weight is increased proportionally; Divide the average value by the preset error reference value to obtain a normalized adjustment ratio; Multiply the directional coefficient by the normalized adjustment ratio to obtain the final adjustment of the parameter weight, ensuring that the adjustment amplitude is linearly related to the degree of deviation; S4024, perform smoothing constraints to limit the adjustment range of a single parameter weight to no more than 10% of the original weight. If the calculated value exceeds the range, it is truncated according to the threshold to ensure model stability; S403, use the historical assimilated data to train a lightweight neural network, input the adjusted difference in S3023 into the network, predict the short-term meteorological element change trend, and feed back the prediction results to the weight allocation module of S200 to update the comprehensive weight matrix.

8. The multi-source heterogeneous meteorological data assimilation and fusion method according to claim 7 is characterized by: S403 is specifically: S4031, extract the difference adjusted by S3023, the generated comprehensive weight matrix and the corresponding timestamp from the historical assimilation data, and generate a time series training sample with a label according to the sliding time window mechanism, where the label is the actual change value of the meteorological element in the next 1 hour; S4032. Construct a lightweight neural network, including an input layer, two convolutional layers, a single-layer long short-term memory unit, and a fully connected output layer, where the convolution kernel size is set to 3×3 and the step size matches the meteorological grid resolution; S4033, using an adaptive learning rate mechanism, with the initial learning rate set to 0.

001. If the validation set loss does not decrease for three consecutive training cycles, the learning rate is decayed to 50% of the original value. At the same time, an early stopping mechanism is introduced to prevent overfitting. S4034, input the adjusted difference value of S3023 at the current moment into the trained network, output the forecast value of the meteorological element change trend in the next hour, and combine the forecast value with the real-time stability index of the data source in S200 to generate a weight correction factor; S4035. Based on the weight correction factors generated in S4034, the comprehensive weight matrix is ​​dynamically adjusted. Specifically, the weights of data sources with high consistency between the forecast trend and the real-time data are increased, and the weights of data sources with low consistency are decreased. The updated weight matrix is ​​directly applied to the next data fusion.

Images