A revised calculation method based on the spatiotemporal neighborhood method for near-surface air temperature and wind speed at grid points

Abstract

The present invention discloses a method for correcting and calculating the near-surface temperature and wind speed of grid points using the spatiotemporal neighborhood method. The method comprises the following steps: obtaining a historical sample data set of the forecast quantity and its corresponding forecast factor; selecting historical samples similar to the current forecast through temporal and spatial classification and preliminary screening; calculating the normalized value of each forecast factor to eliminate the dimensional differences between the factors; calculating the spatiotemporal distance between the current forecast and the historical sample, selecting several historical samples with the smallest distance, and performing weighted averaging based on the analysis values ​​of these samples to generate the final corrected measurement value; testing the corrected measurement results, and optimizing the weight configuration model; in extreme weather conditions, using the seven-day deviation correction method instead of the spatiotemporal neighborhood correction. The method can effectively improve the accuracy of the forecast of near-surface temperature and wind speed in numerical forecasts, reduce the amount of calculation, and effectively improve the forecast and calculation accuracy of near-surface grid temperature and wind speed.

Description

Technical Field

[0001] The present invention belongs to the field of space-time neighborhood method correction and measurement, and specifically relates to a space-time neighborhood method correction and measurement method based on grid point near-ground air temperature and wind speed. Background Art

[0002] The spatiotemporal local correction and measurement of the near-surface air temperature and wind speed can be divided into station correction and measurement and grid correction and measurement. The present invention is specifically aimed at the grid correction and measurement technology of the near-surface air temperature and wind speed.

[0003] Due to the existence of initial value errors and the defects of the numerical model itself, its description of the atmospheric chaotic system inevitably has errors, and its ability to predict meteorological elements near the surface layer is poor. Under the premise of obtaining a high-precision grid surface temperature and wind speed analysis field, based on the forecast results of the numerical model, the grid forecast spatiotemporal method is used to correct the measurement method, which can effectively improve the forecast accuracy of near-surface temperature and wind speed. Its advantage is that it does not require a large amount of computing resources and can effectively improve the forecast accuracy of near-surface temperature and wind speed; its disadvantage is that it must have relatively reliable grid analysis field data as the true value to achieve it. Summary of the invention

[0004] The purpose of the present invention is to provide a spatiotemporal neighborhood method for correcting and calculating the near-surface air temperature and wind speed at grid points, so as to solve the technical problems of large calculation amount and poor forecasting accuracy in the existing schemes.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for correcting and calculating near-surface air temperature and wind speed at grid points using a spatiotemporal neighborhood method, characterized in that the method comprises the following steps:

[0007] Obtaining historical sample data sets including forecast quantities and several forecast factors;

[0008] According to the distribution of the forecast quantity or the forecast factor with the largest contribution rate in the whole time domain, the historical samples are classified and preliminarily screened in time, and the historical samples with the time domain distribution close to the forecast quantity or the forecast factor with the largest contribution rate are selected;

[0009] According to the spatial distribution of the forecast amount or the forecast factor with the largest contribution rate, the historical samples are spatially classified and preliminarily screened to select the historical samples that are close to the required forecast amount or the forecast factor with the largest contribution rate in spatial distribution;

[0010] Calculate the temporal and spatial distances between the same historical samples in the initial screening and the current forecast, select the n historical forecasts with the smallest values, and obtain the corresponding grid point analysis values; perform weighted average on the n historical moment analysis values ​​to obtain the spatiotemporal neighborhood method corrected measurement value of the grid point; the calculation formula for the temporal and spatial distance is:

[0011] ,

[0012] Among them, f(Y T ,y T ) represents the current forecast Y with a forecast validity of T in the l-dimensional space constructed by the forecast factors T and historical forecast T The distance between them; l is the number of selected prediction factors, w i is the weight of the i-th factor, r(y,x i ) is the forecast value y and the original value x of the i-th forecast factor i The statistical correlation coefficient of all historical samples between them; A is the total number of historical samples that have been initially screened, y ave is the average value of the forecast quantity corresponding to all historical samples after the initial screening, x i,ave is the average value of the i-th prediction factor of all historical samples that have been preliminarily screened, ω is the time window; j, k represent the grid longitude and latitude identification numbers respectively, X e,i,T+ω,j,k is the normalized value of the current forecast of the i-th forecast factor at the grid point (j, k) at time T+ω, x e,i,T+ω,j,k is the normalized historical forecast value of the i-th prediction factor at the grid point (j, k) at time T+ω;

[0013] The corrected measurement results are tested, and the weight configuration with the smallest root mean square error is used as the optimal weight to establish the model;

[0014] After normalizing with the current corresponding forecast factors, the established optimal weight configuration model is called to obtain the spatiotemporal neighborhood method corrected estimated values ​​for the current forecast.

[0015] Specifically, the temporal classification includes classifying historical samples according to the numerical distribution of the forecast amount or the forecast factor with the largest contribution rate in the entire time domain, or according to the severity of the numerical change of the forecast amount in the first time window.

[0016] Specifically, the preliminary screening in time includes calculating the difference between the forecast values ​​of the first and last moments of each forecast factor of the current and historical samples within the second time window corresponding to the forecast time, constructing the tendency vector in the multidimensional vector space of the forecast factor with the difference of each forecast factor, and selecting the historical samples whose absolute value of the difference between the constructed current tendency vector and the tendency vector of the historical samples is lower than the first threshold. The first threshold is determined according to the distribution of the historical sample set.

[0017] Specifically, the spatial classification includes classifying the historical samples according to the type of land feature, average elevation, maximum elevation difference and / or vegetation coverage.

[0018] Specifically, the spatial preliminary screening includes selecting historical samples of the same type from the sampling area whose difference with the current terrain index is less than a second threshold value. The second threshold value is determined according to the terrain distribution of the revised measurement area.

[0019] Specifically, the calculation formula of the terrain index Φ is: , δ is the feature type parameter, H ave is the average elevation, ΔH is the maximum elevation difference, q is the vegetation cover parameter, and a, b, c, and d are empirical parameters.

[0020] Specifically, the first time window is larger than the second time window, and the time window is set to an integer multiple of the forecast time validity.

[0021] Specifically, the weighted average of the analysis values ​​at n historical moments is performed to obtain the corrected estimated value of the grid point using the space-time neighborhood method, which is:

[0022] , where f(Y T ,y i,T ) is the distance function value between the nearest historical forecast and the current forecast in the i-th time and space, O T,j,k,i is the analysis value at the grid point (j, k) corresponding to the forecast at the time T closest to the i-th time and space, Y T,j,k,ND-ST It is the corrected value calculated by the space-time locality method at the grid point (j, k) at time T.

[0023] Specifically, under extreme weather conditions, if the historical samples are difficult to match with the current forecast, the seven-day deviation correction method is used instead of the spatiotemporal correction, which specifically includes adjusting the forecast value according to the statistical deviation between the nearest seven-day model forecast and the analysis field.

[0024] Specifically, the normalization process of the prediction factor is processed by the following formula: , where x e is the normalized prediction factor value, x is the original prediction factor value, and x max is the maximum value of the cumulative sample of the predictor, x min The minimum value of the cumulative samples of this predictor.

[0025] The beneficial effects achieved by the present invention are as follows: the method can effectively improve the accuracy of the forecast of near-surface air temperature and wind speed in numerical forecast, reduce the forecast error of grid point near-surface air temperature and wind speed, effectively improve the measurement accuracy of near-surface grid point air temperature and wind speed, and improve the forecast performance. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0027] Figure 1The present invention is a flowchart of the spatiotemporal neighborhood method for correcting and calculating the near-surface air temperature and wind speed at the grid point. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary operation and maintenance personnel in this field without creative work are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, the present invention is a time-space neighborhood method for correcting and calculating grid point near-surface air temperature and wind speed, and its specific steps include:

[0030] Step 10, select a number of forecast factors according to the forecast quantity, and obtain the historical sample data including the time series of the forecast quantity and the forecast factor data.

[0031] Specifically, taking the forecast of 10 m wind speed as an example, its forecast factors can be selected as: 10 m wind speed, 10 m wind direction, 2 m temperature, and ground air pressure. Taking the forecast of 2 m temperature as an example, its forecast factors can be selected as: 2 m temperature, 10 m wind speed, 2 m specific humidity, and ground air pressure.

[0032] Step 20: Classify and preliminarily screen the historical samples in terms of time. This may include:

[0033] Step 21, according to the distribution of the forecast amount or the forecast factor with the largest contribution rate in the whole time domain, the historical samples are temporally classified, and the classified historical samples whose temporal distribution is close to the required forecast amount or the forecast factor with the largest contribution rate are selected.

[0034] According to the distribution of the forecast quantity or the forecast factor with the largest contribution rate in the whole time domain, the historical samples are classified in time. Specifically, they can be classified according to the numerical size distribution state of the forecast quantity in the whole time domain. Alternatively, they can be classified according to the severity of the numerical change of the forecast quantity in the first time window. Alternatively, they can be classified according to the numerical size distribution state of the forecast quantity in the whole time domain and the severity of the numerical change of the forecast quantity in the first time window. For example, according to the numerical size distribution state of the ground temperature throughout the year, it can be divided into two categories of winter and summer half-year or four categories of four seasons. In addition to the numerical size of the forecast quantity, other classification criteria can also be selected, such as the degree of change of the numerical value, such as the variance or standard deviation of the numerical value. Different classification criteria will result in different classifications. For example, the near-ground wind speed can be divided into multiple categories with different wind speed sizes according to the distribution state of the wind speed analysis field, or the historical samples can be divided into two categories of areas with smaller changes and larger changes throughout the year according to the degree of change of the wind speed. The areas with larger changes can be further classified according to the wind speed size distribution.

[0035] Step 22, in the second time window corresponding to the forecast time, calculate the difference between the forecast values ​​of the first and last moments of each forecast factor of the current and historical samples, construct the tendency vector in the multidimensional vector space of the forecast factor with the difference of each forecast factor, and select the historical sample whose absolute value of the difference between the constructed current tendency vector and the tendency vector of the historical sample is lower than the first threshold. The first threshold is determined according to the distribution of the historical sample set. The first time window and the second time window can be set as needed, and can be set to an integer multiple of the forecast time, and the first time window is larger than the second time window.

[0036] Step 30: spatially classify and preliminarily screen the historical samples. Specifically, it may include:

[0037] Step 31, based on the spatial distribution of the forecast amount or the forecast factor with the largest contribution rate, the historical samples are spatially classified, and the classified historical samples that are close in spatial distribution to the required forecast amount or the forecast factor with the largest contribution rate are selected.

[0038] Spatial classification can be based on land feature type and / or average elevation and / or maximum elevation difference and / or vegetation cover. For example, forest-covered mountains can be divided into small areas, areas with drastic changes in surface elevation can be divided separately, and areas with an altitude exceeding a certain height (such as 3,000 or 6,000 meters, Tibet, and the Himalayas) can be divided separately; different types can be divided according to the different surface vegetation cover. For example, they are all mountainous areas, but forests, grasslands, and cultivated land can be distinguished as different types. Detailed spatial area division can increase the accuracy of grid element correction and measurement.

[0039] Step 32, among the same type of historical samples, select the historical samples of the sampling area whose difference with the current terrain index is less than the second threshold, and further select the historical samples of a small area within the grid point 20×20 in the sampling area for modeling. The second threshold is determined according to the terrain distribution of the revised forecast area. The calculation formula of the terrain index Φ is: , δ is the feature type parameter, H ave is the average elevation, ΔH is the maximum elevation difference, q is the vegetation cover parameter, and a, b, c, and d are empirical parameters. The landform types can be divided into 11 categories, namely cultivated land, forest, grassland, shrubland, wetland, water body, tundra, artificial surface, bare land, glacier, and permanent snow. Different landform type parameters δ are assigned to different landform types. The greater the difference in landform types, the greater the difference in landform type parameters δ.

[0040] The most important thing about the revised measurement method of the present invention is to find the most similar historical samples. The classification and preliminary screening in time and space of the present invention reduces the amount of calculation and improves the accuracy to a certain extent. When searching for the closest historical samples in subsequent modeling, only the historical samples after preliminary screening in time and space are searched.

[0041] Modeling is performed on small areas that are finely divided in space and historical samples that are preliminarily classified in time. The most similar historical samples are found according to the same time category and the same surface type after initial screening. The forecast quantity grid analysis values ​​corresponding to a fixed number (such as 20) of historical samples are weighted averaged to give the final deterministic corrected measurement value, providing a more accurate corrected measurement technology. At the same time, the forecast quantity analysis values ​​corresponding to a fixed number (such as 20) of historical samples can also provide a probabilistic forecast of the ensemble forecast concept.

[0042] Step 40, the historical sample data after initial screening with the same reporting start time and the same forecast validity period are calculated using the formula: , calculate the normalized value of each prediction factor, de-dimensionalize it, and thus eliminate the impact of the modeled prediction factors due to their different units. e is the normalized prediction factor value, x is the original prediction factor value, and x max is the maximum value of the cumulative sample of the predictor, x min is the minimum value of the cumulative samples of the prediction factor. The prediction factor after normalization is dimensionless and its value range is [0,1]. This avoids the large difference in the initial contribution to the subsequent distance function due to the different value units of the prediction factor itself.

[0043] Step 50, calculate the temporal and spatial distance between the historical samples of the same time and the current forecast, select the n (such as n=20) historical forecasts with the smallest value and obtain the corresponding grid analysis value. The calculation formula of the temporal and spatial distance is:

[0044]

[0045] Among them, f(Y T ,y T ) represents the current forecast Y with a forecast validity of T in the l-dimensional space constructed by the forecast factors T and historical forecast T The smaller the value, the closer they are in time and space; l is the number of selected prediction factors, w i is the weight of the i-th factor, r(y,x i ) is the forecast quantity y and the original value x of the i-th forecast factor i The statistical correlation coefficient of all historical samples between the two. The more relevant it is to the forecast quantity, that is, the larger the absolute value of the correlation coefficient, the greater its contribution to calculating the distance between the current forecast and the historical forecast. This coefficient is the same in the modeling stage and the forecasting stage. As the samples are increased or updated, the model is updated, and its value is also updated accordingly. A is the total number of historical samples that have been initially screened, and y ave is the average value of the forecast quantity corresponding to all historical samples after the initial screening, x i,ave is the average value of the i-th prediction factor of all historical samples that have been initially screened, ω is the time window, indicating that the forecast times [T-τ, T+τ] that are close in time participate in the distance calculation in the space-time domain; the time weight function is introduced into the formula to determine the weight of each forecast time in the time window. The closer the time is, the greater the weight is. By selecting this weight function, the accuracy of space-time distance calculation can be further improved, thereby improving the efficiency and accuracy of the method; j, k represent the grid longitude and latitude identification numbers respectively, X e,i,T+ω,j,k is the normalized value of the current forecast of the i-th forecast factor at the grid point (j, k) at time T+ω, x e,i,T+ω,j,k is the i-th predictor The normalized value of the historical forecast at the grid point (j, k) at the moment. The distance between all grid points in the grid area and the historical forecast is calculated. The spatial distance is calculated only between samples of the same time type, and the time type is determined by the distribution state of the analysis value of the forecast quantity. Therefore, the nearest n (such as n=20) historical forecast samples determined by the distance function value calculated by the formula are all "neighboring" samples in time and space.

[0046] By examining the 2 m temperature and 10 m wind speed RMAPS-RISE analysis fields and the forecast lag correlation coefficients of the numerical forecast model (CMA-BJ), it can be seen that the correlation is best at the same time, and the larger the time interval, the smaller the correlation; the temporal proximity in modeling and forecasting is reflected in: only looking for samples of spatial proximity among samples of the same time type that have passed the initial screening; the weight of the forecast time involved in modeling is related to the time difference between it and the required forecast time, and the larger the time difference between it and the required forecast time, the smaller the weight.

[0047] Step 60, weighted average of historical forecast analysis values, that is, to obtain the time-space neighborhood method corrected estimated value of the grid point. Specifically, from all the time-space distances of all the historical samples of the same time and space that have been preliminarily screened and the current forecast calculation, select the n (such as n = 20, which is equivalent to selecting the n points closest in time and space in l-dimensional space) historical forecast corresponding observation values ​​with the smallest value, and calculate according to the following formula:

[0048] , where f(Y T ,y i,T ) is the distance function value between the nearest historical forecast and the current forecast in the i-th time and space, O T,j,k,i is the analysis value at the grid point (j, k) corresponding to the forecast at the time T closest to the i-th time and space, Y T,j,k,ND-ST It is the corrected value calculated by the space-time locality method at the grid point (j, k) at time T.

[0049] Since samples are classified in time and space, small areas are modeled according to surface types, and small area forecasts are combined into a full-area grid forecast field. In this way, the computing resources required for modeling are greatly reduced. Moreover, since the classification is carried out in space, full-area grid forecasts can be achieved with only a few models, that is, covering all terrains. Moreover, since small-area modeling and small-area forecasting are carried out, the forecast effect can be better guaranteed.

[0050] Step 70, checking the test set spatiotemporal neighborhood method correction measurement results, taking the weight configuration with the smallest root mean square error as the optimal weight, and establishing the model;

[0051] Step 80, after normalizing with the current corresponding forecast factors, call the established optimal weight configuration model to obtain the spatiotemporal neighborhood method corrected estimated value for the current forecast.

[0052] In the event of extreme weather that has not appeared in the historical sample set, making it difficult for the historical sample to match the current forecast, the seven-day deviation correction is used to replace the spatiotemporal correction. Specifically, the full-field average of the forecast value of the model forecast is used to determine whether extreme weather will occur. If so, the statistical deviation between the nearest seven-day model forecast and the analysis field is added to the model grid forecast to form the final grid correction measurement result. Otherwise, the spatiotemporal correction is still used.

[0053] The following is a specific application example of this solution, based on the hourly near-surface forecast element data with a horizontal resolution of 3km of the Beijing Meteorological Bureau's operational numerical forecast model (CMA-BJ), and the analysis field is the Beijing-Tianjin-Hebei near-surface temperature and wind speed grid data provided by RMAPS-RISE. Based on the CMA-BJ numerical model forecast data with an hourly forecast of 1-72h from 03UTC (universal time) from January 1, 2021 to September 22, 2022, and the near-surface temperature and wind speed analysis field data of RMAPS-RISE, the model is modeled with data from January 1, 2021 to July 31, 2022, and the near-surface temperature and wind speed from August 1, 2022 to September 22, 2022 are corrected and calculated to test and compare their forecast effects.

[0054] First, the sample time is classified. For the convenience of discussion, the sample time is simply divided into the winter half year (November 1st - April 30th of the following year) and the summer half year (May 1st - October 31st). Secondly, since the overall terrain distribution of Beijing-Tianjin-Hebei is relatively clear and there is no drastic change in terrain, it is roughly divided into four categories according to the surface type: plains, mountains, oceans, and the junction of plains and mountains. Alternatively, it can be further refined, such as forests, grasslands, and cultivated land. Further refinement can slightly improve the forecast accuracy on this basis. In order to reduce the computing resources required for modeling of this method, in this application example, it is classified according to the surface type and further divided into small areas. The Beijing-Tianjin-Hebei region can be divided into: 51 mountain areas, 37 plain areas, 9 ocean areas, and 7 mountain-plain mixed areas. A total of 104 areas of four surface types are used for initial screening, modeling and forecasting.

[0055] The numerical model forecast adopts the forecast results of the Beijing Meteorological Bureau's operational numerical forecast model from 03UTC for 1-72h with a horizontal resolution (grid spacing) of 3km hourly. The 2m temperature and 10m wind speed analysis fields use the Beijing-Tianjin-Hebei grid data corresponding to the numerical model from Ruitu-Ruisi. The modeling data is from January 1, 2021 to July 31, 2022. Taking January 1, 2021 as an example, the data corresponding to the time is from 03UTC on January 1, 2021 to 03UTC on January 4, 2021. In the modeling sample set (577 days), the last 100 days of samples are selected as the test set, and the minimum root mean square error of the forecast test set is used as the optimal judgment criterion. A small area is selected as a representative according to the surface type for modeling, and the optimal weights of each forecast factor at each forecast time for each surface type are obtained, that is, the modeling of four types of surface types is completed. The spatiotemporal neighborhood method correction and calculation technology of the present invention is adopted, and the prediction factor grid value of the CMA-BJ numerical prediction model with a horizontal resolution of 3km starting from 03UTC on August 1, 2022 to September 22, 2022 is used. Through the established classification model, the near-surface temperature and wind speed with a resolution of 3km in the Beijing-Tianjin-Hebei region are corrected and calculated.

[0056] A statistical comparative analysis of the spatiotemporal neighborhood method correction measurement effects was conducted based on a total of 36,504 grid points (169 (longitude) × 216 (latitude) in the Beijing-Tianjin-Hebei region.

[0057] (1) Statistics by forecast time

[0058] Statistics are made by forecast time, that is, all grid forecast data of Beijing-Tianjin-Hebei in the forecast sample set at the same forecast time are compared with the Ruitu-Ruisi analysis field data, and then the overall forecast effect is averaged over all times. Before and after the correction by the spatiotemporal domain method, the 2m temperature root mean square error is reduced by 0.42℃, and the absolute value of the forecast deviation is smaller after correction, of which 98.6% of the forecast times are improved; the 10m wind speed root mean square error is reduced by 0.92m / s, and the absolute value of the forecast deviation is smaller after correction, of which 100% of the forecast times are improved. The forecast accuracy is significantly improved compared with the CMA-BJ numerical model forecast.

[0059] (2) Statistics by grid points

[0060] By grid point statistics, the forecast data of all times of each grid point forecast sample set is compared with the Ruitu-Ruisi analysis field, and then the overall forecast effect is obtained by averaging all grid points. Before and after the correction of the spatiotemporal neighborhood method, the root mean square error of 2m temperature decreased by 0.38℃, of which 84.3% of the grid points had improved forecast effects, and the absolute value of forecast deviations was smaller after correction; the root mean square error of 10m wind speed forecast decreased by 1.05m / s, of which 93.9% of the grid points had improved forecast effects, and the absolute value of forecast deviations was smaller after correction.

[0061] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative.

[0062] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic features of the present invention.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary operation and maintenance personnel in the field should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for calculating the temperature and wind speed near the ground using the spatiotemporal neighborhood correction method at grid points, characterized in that: The method comprises the following steps: According to the forecast quantity, select several forecast factors, and obtain the historical sample data including the time series of forecast quantity and forecast factor data; According to the distribution of the forecasting factor with the largest forecast amount or contribution rate in the whole time domain, the historical samples are classified and preliminarily screened in time. The preliminarily screened in time includes calculating the difference between the forecast values ​​of the first and last moments of each forecasting factor of the current and historical samples in the second time window corresponding to the forecasting time, constructing the tendency vector in the multidimensional vector space of the forecasting factor with the difference of each forecasting factor, and selecting the historical samples whose absolute value of the difference between the constructed current tendency vector and the tendency vector of the historical samples is lower than the first threshold value determined according to the distribution of the historical sample set; According to the spatial distribution of the forecasting factor with the largest forecast amount or contribution rate, the historical samples are spatially classified and preliminarily screened. The spatial preliminarily screening includes selecting historical samples of the sampling area whose difference with the current terrain index is less than the second threshold determined according to the revised terrain distribution of the measured area from the same historical samples. The calculation formula of the terrain index Φ is: , δ is the feature type parameter, H ave is the average elevation, ΔH is the maximum elevation difference, q is the vegetation cover parameter, and a, b, c, and d are empirical parameters; Calculate the temporal and spatial distances between the same historical samples in the initial screening and the current forecast, select the n historical forecasts with the smallest values, and obtain the corresponding grid point analysis values; perform weighted average on the n historical moment analysis values ​​to obtain the spatiotemporal neighborhood method corrected measurement value of the grid point; the calculation formula for the temporal and spatial distance is: ， Among them, f(Y T ,y T ) represents the current forecast Y with a forecast validity of T in the l-dimensional space constructed by the forecast factors T and historical forecast T The distance between them; l is the number of selected prediction factors, w i is the weight of the i-th factor, r(y,x i ) is the forecast value y and the original value x of the i-th forecast factor i The statistical correlation coefficient of all historical samples between them; A is the total number of historical samples that have been initially screened, y ave is the average value of the forecast quantity corresponding to all historical samples after the initial screening, x i,ave is the average value of the i-th prediction factor of all historical samples that have been preliminarily screened, ω is the time window; j, k represent the grid longitude and latitude identification numbers respectively, X e,i,T+ω,j,k is the normalized value of the current forecast of the i-th forecast factor at the grid point (j, k) at time T+ω, x e,i,T+ω,j,k is the normalized historical forecast value of the i-th prediction factor at the grid point (j, k) at time T+ω; The corrected measurement results are tested, and the weight configuration with the smallest root mean square error is used as the optimal weight to establish the model; After normalizing with the current corresponding forecast factors, the established optimal weight configuration model is called to obtain the spatiotemporal neighborhood method corrected estimated values ​​for the current forecast.

2. The method according to claim 1, characterized in that The temporal classification includes classifying the historical samples according to the numerical distribution of the forecast amount or the forecast factor with the largest contribution rate in the entire time domain, or according to the severity of the numerical change of the forecast amount in the first time window.

3. The method according to claim 1, characterized in that The spatial classification includes classifying the historical samples according to the type of land feature, average elevation, maximum elevation difference and / or vegetation coverage.

4. The method according to claim 2, characterized in that: The first time window is larger than the second time window, and the time window is set to an integer multiple of the forecast time validity.

5. The method according to claim 1, characterized in that The weighted average of the analysis values ​​at n historical moments is used to obtain the corrected estimated value of the grid point by the space-time neighborhood method, which is: , where f(Y T ,y i,T ) is the distance function value between the nearest historical forecast and the current forecast in the i-th time and space, O T,j,k,i is the analysis value at the grid point (j, k) corresponding to the forecast at the time T closest to the i-th time and space, Y T,j,k,ND-ST It is the corrected value calculated by the space-time locality method at the grid point (j, k) at time T.

6. The method according to claim 1, characterized in that Under extreme weather conditions, if the historical samples are difficult to match with the current forecast, the seven-day deviation correction method is used instead of the spatiotemporal correction, which specifically includes adjusting the forecast value according to the statistical deviation between the nearest seven-day model forecast and the analysis field.

7. The method according to claim 1, characterized in that The normalization process of the prediction factor is processed by the following formula: , where x e is the normalized prediction factor value, x is the original prediction factor value, and x max is the maximum value of the cumulative sample of the predictor, x min The minimum value of the cumulative samples of this predictor.

Images